It is worth mentioning, that there are two types of ephemeral tables:
one holds rows as they are inserted (all rows turn out to be distinct);
another adds ROWID column in order to hold equal rows.  Inasmuch as
original ROWID mechanism is to be removed, the purpose of this patch is
to emulate ROWID behavior for ephemeral tables.

 - Substituted all ephemeral table opcodes which don't rely on SQLite's
   ROWID with Tarantool ones.
 - Implemeted opcode OP_NextIdEphemeral to emulate behavior of ROWID
   feature.  In current implemetation ROWID is not hidden PK index, but
   just another one column to keep in table equal tuples.
 - Substituted old ephemeral tables with ROWID for INSERT INTO ...
   SELECT ... FROM statement. It utilizes ephemeral tables as
   intermediate holder for data from SELECT
   (before it is inserted to table).

Part of #2680

 - Completed implementing tarantoolSqlite3Ephemeral* API.  Added
   functions: Last, Next, Previous, Delete, Unpack.
 - Added support of ephemeral tables for IN expression.
 - Added support of ephemeral tables for sub-select queries.
 - Fixed tests, which rely on the name of ephemeral creation opcode.

Part of #2680

 - Added intermediate functions between SQL and Tarantool,
   tarantoolSqlite3Ephemeral* in order to create and use Tarantool's
   ephemeral tables. Also, for this reason added BTCF_TEphemeral flag to
   BtCursor instead of BTCF_TaCursor to call appropriate functions.
 - Added struct space* to SQL cursor, which allows to delete ephemeral
   tables.  (since ephemeral tables don't have space_id and it is
   impossible to find them by traditional space_id lookup). Moreover, it
   enables us to avoid using box API inasmuch as we can directly call
   space_execute_* functions having struct space*.
 - As the first stage of implemenation, ephemeral tables were
   successfully replaced for DELETE FROM ... WHERE statement. It utilized
   ephemeral space as intermediate holder where all tuples to be removed
   are placed before actual deletion.

Part of #2680

Ephemeral tables shouldn't be involved in any transaction routine,
since they are used only for internal needs. Moreover, ephemeral spaces
can be created and destroyed within one transaction, which may lead to
incorrect behaviour. Thus, ephemeral replace and delete don't take txn
as an argument.

Furthermore, delete operation should work on tuples with nulls in primary key.
As a rule, it is considered to be bad practice, due to ambiguity when
deleting tuple from space with several tuples containing nulls in PK.
On the other hand, ephemeral spaces are used only for internal needs,
so if it is guaranteed that no such situation occur, it is OK to allow
insertion nulls in PK. Hence, function which implements deletion on ephemeral
tables, doesn't check tuple on null fields.

Currently, supported only by memtx engine.

By now, replace function takes new tuple and old tuple as arguments, instead
of single txn_stmt. It has been done in order to avoid abusing txn_stmt:
the only usage was extracting tuples from it.
As a result, this function can be used by ephemeral tables
without any patching.

Ephemeral spaces are invisible via public API and they are not
persistent. They are needed solely to do some transient calculations.

To create an ephemeral space, use space_new_ephemeral() function,
which takes a space definition and a list of index definitions as
arguments. A space created with this function may be accessed and
modified via internal API (space_execute_replace(), index_get(),
etc).

Currently, only the memtx engine supports ephemeral spaces.

Closes #2776

Add forgotten box/sql/parse.c source which is being
regenerated in -DSQL_MAINTAINER_MODE=1 only.

Macro SQLITE_DEBUG was set unconditionally, so both release build
modes had debugging enabled, including all asserts() and TESTCASE().
Move this macro definition under check if mode is Debug and fix all
warning which arised after disabling asserts().
Most of them are unused variables.

Since asserts were always disbled in src/box/sql.c, number
of issues were hidden. This patch enables asserts for mentioned
file by disabling set of NDEBUG in box/src/sqliteInt.h.
Also this patch fixes incorrect assertions about strlen() and
assertion that fieldno should always be less than format->field_count.
This assert is not always hold, e.g. for _schema space where max_id
is stored behind formatted fields.

Closes #3057

Add forgotten in previous commit regenerated box/sql/parse.c

Currently, only ALTER TABLE RENAME is supported in SQL. Ban
ALTER TABLE ADD COLUMN in the grammar until it is supported.
Corresponding issue submitted: #3075

Closes #2607

There're 3 places where FreeBSD specific routines
contained (for compatibility) unused parameters.
Mark them w/ unused attribute.

Closes #3037

SQLITE_ENABLE_UNLOCK_NOTIFY assumed to be always disabled.
SQLITE_OMIT_SHARED_CACHE assumed to be always enabled.

Deleted files
-------------

src/box/sql/notify.c
src/box/sql/btmutex.c

Deleted non-static functions
----------------------------

// btmutex.c
sqlite3BtreeEnter()
sqlite3BtreeLeave()
sqlite3BtreeHoldsMutex()
sqlite3BtreeEnterAll()
sqlite3BtreeLeaveAll()
sqlite3BtreeHoldsAllMutexes()
sqlite3SchemaMutexHeld()
sqlite3BtreeEnterCursor()
sqlite3BtreeLeaveCursor()

// btree.c
sqlite3BtreeLockTable()
sqlite3BtreeSharable()
sqlite3BtreeConnectionCount()

// notify.c
sqlite3_unlock_notify()
sqlite3ConnectionBlocked()
sqlite3ConnectionUnlocked()
sqlite3ConnectionClosed()

// vdbeaux.c
sqlite3VdbeEnter()
sqlite3VdbeLeave()

Deleted structure fields
------------------------

// btreeInt.h
struct Btree.wantToLock
struct Btree.lock
struct BtShared.nRef
struct BtShared.pNext
struct BtShared.pLock
struct BtShared.pWriter

// sqliteInt.h
struct sqlite3.pBlockingConnection
struct sqlite3.pUnlockConnection
struct sqlite3.pUnlockArg
struct sqlite3.xUnlockNotify
struct sqlite3.pNextBlocked

Deleted asserts
---------------

// common
assert(sqlite3BtreeHoldsAllMutexes(...))
assert(sqlite3SchemaMutexHeld(...))
assert(sqlite3BtreeHoldsMutex(...))

// btree.c
assert(... || sqlite3BtreeConnectionCount(...) > 0)
assert(p->wantToLock == 0 && p->locked == 0)
assert(hasSharedCacheTableLock(...))
assert(... || !hasReadConflicts(...))

Part of #2931

  There are some cases where a number of similar tests should be performed.
In this scenario, we had to write a lot of redundant symbols.

  The new method allows one to write the minimal number of code necessary for
tests.

This comit introduces a number of changes:

1. move a transaction state to fiber local struct

  This is important because the `sqlite3` is a shared structure and it was
used to store data related to the transaction. However it was working because
yield is called only on commit and it garanteed unique access. (With possible
effects on ddl.)

NOTE: `nDeferredCons` and `nDeferredImmCons` are stored in vdbe during vdbe
execution and moved to sql_txn when it needs to be saved until execution of the
next vdbe in the same transaction.

2. support savepoints

2.1. support abort (anonymous savepoints)

  Abort mechanism was simplified. Instead of storing track of all
savepoints without name, this commit introduces `anonymous_savepoint`.
  `anonymous_savepoint` is a structure which is stored in Vdbe and represents
the state of database on the beginning of the current statement. Tarantool
disallow multistatement, so a vdbe can have one statement max. That is why
having one savepoint is enough to perform abort.

2.2. named savepoints

Key points:
  - It uses Tarantool's savepoints
  - It allocates savepoints on the "region" (they are destroyed automatically)
  - There are some crutches around ddl (ddl should not be placed into a
transaction)

Closes #2989 #2931 #2964

The header is useless for Tarantool.

Fixes #2944.

There were dozen of test cases which still used ROWID feature.
Patch updates such cases.

Fixes #2884.

Rowid is removed right now from Tarantool SQL. That means that table
must have primary key. However, table defined w/o PK field caused
Tarantool to a crash instead of raising an error message complaining
missing PRIMARY KEY. Fix this and add a regression test

Closes #2929

Assign is_nullable property for each column in order to fix all sql
tests after reverting MP_NIL comparators in previous commit.

Closes #3009

This reverts commit a2f574e3.

After TableLock opcode removal, need to re-generate
opcode definitions file.

Removed OP_ParseSchema, OP_TableLock.
All other dead code connected to locking of open tables (for both
reading and writing) removed as well.
Remove auth.c file, it is not used anymore.

Few opcodes like OpenRead/OpenWrite exepcted 3rd argument to be DB
descriptor. Now there's a single DB backend and there's no need
for any descriptors. Remove from VDBE, fix code generator, update
opcode descriptions.

Closes #2855

  Make collation tokens obey ANSI ISO standard rules for identifiers.

Related to #2818

Closes #2653, #2121

  Thsis patch do two main changes:
  1. Delete SQLite's coll structure & replace it with Tarantool's
  2. Delere SQLite's create coll mechanism

  Tarantool can not easily support create collation mechanism of SQLite,
because collations must:
  - be persistent
  - be the same in tuple comparator and SQL Mem comparator
  Although new collations are limited, suggested ICU collations cover the
main set of use-cases. ICU supports:
  - Different locales & different strengths & other options...
  - Nimeruc stringths comparation.

  That is why we limit collation api at this point.

Part of #2121

  collation_list pragma should change as the way collations work changes.
To enumerate all collations one have to iterate space for now.

  This feature is required by SQL, because searching collations by name is
common operation there.

  Is implemented by creating extra hash table.

This is the second attempt to remove encodings.
At this point Tarantool SQL supports only UTF-8 encoding.

Part of #2295

Part of #2121

Deletion of rows from table (which is part of DROP TABLE procedure) is
wrapped in transaction now. If any Foreign Key violations occur, then
rollback transaction and halt execution of statement. Otherwise, commit
changes and continue DROP TABLE routine. This logic is implemented by
OP_FkCheckCommit opcode.
Added tests for this case.

Closes #2953

  There were analogies of those functions in Tarantool. This commit makes
SQL reuse them.
  All crutches (undef) are deleted.

Added new opcode OP_RenameTable, which implements
SQL statment: ALTER TABLE old_name RENAME TO new_name;
The main idea is to replace corresponding tuple in _space.
New tuple will contain new name and new SQL statement, which
creates this table. If it is parent table for foreign key,
update SQL statement of child table. Then, remove old name
from hash table and by calling callback function from VDBE,
update database schema with new table name. After all,
in the same way update triggers' statements in _trigger table,
if any exist.

Closes #2204

Closes #2506