box: don't delete inprogress files during normal GC

*.inprogress files shouldn't be deleted during normal GC, because they
may be in use. E.g. GC may happen to run while WAL is rotating xlog;
if GC removes the transient xlog.inprogress file created by WAL (see
xdir_create_xlog), WAL will fail to rename it and abort the current
transaction.

Initially, inprogress files were removed only after recovery. For this
reason, xdir_collect_inprogress, which is used for deleting inprogress
files, accesses FS directly, without COIO. This was changed by commit
5aa243de ("recovery: build secondary
index in hot standby mode") which moved xdir_collect_inprogress from
memtx_engine_end_recovery to memtx_engine_collect_garbage so that a hot
standby instance doesn't delete inprogress files of the master instance
by mistake.

To fix this issue, let's move xdir_collect_inprogress back where it
belongs, to engine_end_recovery, and introduce a new callback for memtx
to build its secondary keys before entering the hot standby mode -
engine_begin_hot_standby.

Let's also remove engine_collect_garbage from gc_init, which was added
there by the aforementioned commit, probably to make tests pass.

The bug was reported by the vinyl/deferred_delete test (#5089).

Closes #6554