* update comments
* add a test case for altering a primary key on the fly
* rename AddIndex to CreateIndex
* factor out common code into a function

MoveIndex operation is used to move an existing index from the old space
to the new one. Semantically it's a no-op.

RebuildIndex is introduced for case when essential index properties are
changed, so it is necessary to drop the old index and create a new one in
its place in the new space.

AlterSpaceOp::prepare() is removed: all checks are moved to
on_replace trigger in _index system space from it.
All checks are done before any alter operation is created.

Necessary for gh-2074 and gh-1796.

Move the check which decides on an alter strategy whenever
a row in _index space is changed, from AlterSpaceOp::preapre()
to on_replace trigger on _index space.

The check chooses between two options: a heavy-weight index rebuild,
invoked when index definition, such as key parts, is changed, vs.
lightweight modify, invoked when index name or minor options are
modified..

Before this patch, index alteration creates a pair of operations
(DropIndex + AddIndex) in all cases, but later
replaces two operations with one at AlterSpaceOp::prepare() phase.

This is bad by several reasons:

- it's done while traversing of a linked list of operations,
  and it changes the list being traversed.

- an order in the list of operations is required for this
to work: drop must precede add.

- needless allocation and deallocation of operations makes the logic
  unnecessarily complex.

Necessary for gh-1796.

Always first create the primary key index in a space. Put
the primary key key def first in the array of key_defs, passed
into tuple_format_new(). This is necessary for gh-1796.

Assert that we can't create a space with secondary key but no primary.

Now drop primary index checks are made in alter triggers after
new space creation. Such an implementation leads to temporary
creation of a space with invalid index set. Fix it and check
the index set before space_new call.

Non destructive swap_index_def function sould be used because memtx_tree
stores a pointer to an index_def used while tree creation. Fixed #2570

Commit dbfd515f ("vinyl: fix crash if snapshot is called while dump
is in progress") introduced a bug that can result in statements inserted
after WAL checkpoint being included in a snapshot. This happens, because
vy_begin_checkpoint() doesn't force rotation of in-memory trees anymore:
it bumps checkpoint_generation, but doesn't touch scheduler->generation,
which is used to trigger in-memory tree rotation.

To fix this issue, this patch zaps scheduler->checkpoint_generation and
makes vy_begin_checkpoint() bump scheduler->generation directly as it
used to. To guarantee dump consistency (the issued fixed by commit
dbfd515f), scheduler->dump_generation is introduced - it defines the
generation of in-memory data that are currently being dumped. The
scheduler won't start dumping newer trees until all trees whose
generation equals dump_generation have been dumped. The counter is only
bumped by the scheduler itself when all old in-memory trees have been
dumped. Together, this guarantees that each dump contains data of the
same generation, i.e. is consistent.

While we are at it, let's also remove vy_scheduler->dump_fifo, the list
of all in-memory trees sorted in the chronological order. The scheduler
uses it to keep track of the oldest in-memory tree, which is needed to
invoke lsregion_gc(). However, since we do not remove indexes from the
dump_heap, as we used to not so long ago, we can use the heap for this.
The only problem is indexes that are currently being dumped are moved
off the top of the heap, but we can detect this case by maintaining a
counter of dump tasks in progress: if dump_task_count is > 0 when a dump
task is completed, we must not call lsregion_gc() irrespective of the
generation of the index at the top of the heap. A good thing about
ridding of vy_scheduler->dump_fifo is that it is a step forward towards
making vy_index independent of vy_scheduler so that it can be moved to a
separate source file.

Closes #2541
Needed for #1906

vy_index->generation equals to the generation of the oldest in-memory
tree, which can be looked up efficiently as vy_index->sealed list is
sorted by generation so let's zap it and add vy_index_generation()
function instead.

Needed for #1906

Including index.h just for the sake of iterator_type, as we do in
vy_run.h and vy_mem.h, is a bit of overkill. Let's move its definition
to a separate source file, iterator_type.h.

 - Replace vy_range->index with key_def.
 - Replace vy_range_iterator->index with vy_range_tree_t.

Needed for #1906

The compact_heap, used by the scheduler to schedule range compaction,
contains all ranges except those that are currently being compacted.
Since the appropriate vy_index object is required to schedule a range
compaction, we have to store a pointer to the index a range belongs to
in vy_range->index. This makes it impossible to move vy_range struct and
its implementation to a separate source file.

To address this, let's rework the scheduler as follows:

 - Make compact_heap store indexes, not ranges. An index is prioritized
   by the greatest compact_priority among its ranges.

 - Add a heap of ranges to each index, prioritized by compact_priority.
   A range is removed from the heap while it's being compacted.

 - Do not remove indexes from dump_heap or compact_heap when a task is
   scheduled (otherwise we could only schedule one compaction per
   index). Instead just update the index position in the heaps.

Needed for #1906

Advancing replica->gc on every status update is inefficient as gc can
only be invoked when we move to the next xlog file. Currently, it's
acceptable, because status is only updated once per second, but there's
no guarantee that it won't be updated say every millisecond in future,
in which case advancing replica->gc on every status update may become
too costly.

So introduce a trigger invoked every time an xlog is closed by
recover_remaining_wals() and use it in relay to send a special gc
message.

ipc_cond_wait() always returns 0, so the body of the loop waiting for
the endpoint to be ready for destruction is only invoked once.

To make sure there is no status message pending in the tx pipe,
relay_cbus_detach() waits on relay->status_cond before proceeding to
relay destruction. The status_cond is signaled by the status message
completion routine (relay_status_update()) handled by cbus on the
relay's side. The problem is by the time we call relay_cbus_detach(),
the cbus loop has been stopped (see relay_subscribe_f()), i.e. there's
no one to process the message that is supposed to signal status_cond.
That means, if there happens to be a status message en route when the
relay is stopped, the relay thread will hang forever.

To fix this issue, let's introduce a new helper function, cbus_flush(),
which blocks the caller until all cbus messages queued on a pipe have
been processed, and use it in relay_cbus_detach() to wait for in-flight
status messages to complete. Apart from source and destination pipes,
this new function takes a callback to be used for processing incoming
cbus messages, so it can be used even if the loop that is supposed to
invoke cbus_process() stopped.

 - Fold in wal dir scan. It's pretty easy to detect if we need to rescan
   wal dir - we do iff the current wal is closed (if it isn't, we need
   to recover it first), so there's no point in keeping it apart.

 - Close the last recovered wal on eof. We don't close it to avoid
   rereading it in case recover_remaining_wals() is called again before
   a new wal is added to wal dir. We can detect this case by checking if
   the signature of the last wal stored in wal dir has increased after
   rescanning the dir.

 - Don't abort recovery and print 'xlog is deleted under our feet'
   message if current wal file is removed. This is pointless, really -
   it's OK to remove an open file in Unix. Besides, the check for
   deleted file is only relevant if wal dir has been rescanned, which is
   only done when we proceed to the next wal, i.e. it doesn't really
   detect anything.

A good side effect of this rework is that now we can invoke garbage
collector right from recovery_close_log().

* use per-index statistics
* remove step_count as it is no longer maintained
* add statistics for txw, mem, and index overall

Old iterator has several problems:

- restoration system is too complex and might cause several
  reads from disk of the same statements.

- applying of upserts works in direct way (squash all upserts
  and apply them to terminal statement) and the code doesn't
  leave a change to optimize it.

Implement iterator for full-key EQ case that fixes problems above.

There is a version member in vy_index that is incremented on each
modification of mem list and range tree. Split it to two members
that correspond to mem list and range tree accordingly.

It is needed for more precise tracking of changes in iterators.

If a vylog record doesn't get flushed to disk due to an error, objects
it refers to (index->key_def, range->begin and range->end) may get
destroyed, resulting in a crash. To avoid that, we must copy those
objects to vylog buffer.

Closes #2532

 - Remove tx and cursor latencies as they are useless - they actually
   account how long a tx/cursor was open, not latencies.
 - Remove vy_stat->get_latency as it doesn't account latency of select,
   besides we now have per index read latency.
 - Remove vy_stat->dumped_statements and dump_total as these statistics
   are reported per index as well.
 - VY_STAT_TX_OPS is currently unused (always 0). Let's use it for
   accounting the total number of statements committed in tx instead of
   vy_stat->write_count.

index.info() is supposed to show index stats, not options.
box.space.<space_name>.index.<index_name>.options looks like
a better place for reporting index options.

Needed for #1662

This patch adds 'latency' field to index.info. It shows the latency of
reads from the index. The latency is computed as 99-percentile of all
delays incurred by vy_read_iterator_next().

Needed for #1662

Replace box.info.vinyl().performance.upsert_{squashed,applied} with per
index index.info().upsert.{squashed,applied}.

Needed for #1662

Add the following counters to index.info:

  lookup                # number of lookups (read iter start)
  get                   # number of statements read (read iter next)
    rows
    bytes
  put                   # number of statements written
    rows
    bytes

Needed for #1662

closes gh-2516

This patch adds the following counters to index.info:

  txw
    count               # number of statements in the TX write set
      rows
      bytes
    iterator
      lookup            # number of lookups in the TX write set
      get               # number of statements returned by the iterator
        rows
        bytes

Needed for #1662

This patch adds the cache section to index.info with the following
counters in it:

  cache
    rows                # number of tuples in the cache
    bytes               # cache memory size
    lookup              # lookups in the cache
    get                 # reads from the cache
      rows
      bytes
    put                 # write to the cache
      rows
      bytes
    invalidate          # overwrites in the cache
      rows
      bytes
    evict               # evictions due to memory quota
      rows
      bytes

Needed for #1662

Using vy_quota, which was implemented to support watermarking,
throttling, timeouts, for accounting cached tuples is an overkill.
Replace it with mem_used and mem_quota counters.

This patch adds the following counters to the disk section index.info:

  dump                  # dump statistics:
    count               #   number of invocations
    in                  #   number of input statements
      rows
      bytes
    out                 #   number of output statements
      rows
      bytes

  compact               # compaction statistics:
    count               #   number of invocations
    in                  #   number of input statements
      rows
      bytes
    out                 #   number of output statements
      rows
      bytes

Needed for #1662

Replace box.info.vinyl().performance.iterator.{run,mem} global counters
with the following per index counters:

  memory
    iterator
      lookup            # number of lookups in the memory tree
      get               # number of statements returned by mem iterator
        rows
        bytes

  disk
    iterator
      lookup            # number of lookups in the page index
      get               # number of statements returned by run iterator
        rows
        bytes

      bloom             # number of times bloom filter
        hit             #   allowed to avoid a disk read
        miss            #   failed to prevent a disk read

      read              # number of statements actually read from disk
        rows
        bytes
        bytes_compressed
        pages

Needed for #1662

It's useless - it only checks that certain counters are present, but
doesn't give a damn about what they show. At the same time, I have to
update its output every time I modify vinyl statistics, which I'm doing
pretty often these days. So let's ditch it - I'll rewrite it from
scratch after I'm done reworking vinyl statistics.

gh-2520: add comments explaining the test case for gh-2520 (upsert
caching).