MDL WHEN LOCK=SHARED".

ALTER TABLE executed using in-place algorithm could trigger an
InnoDB assertion if during prepare phase of ALTER TABLE, the
table being altered happened to be used in more than one connection.

The problem was that for types of in-place ALTER TABLE which
require strong lock during the prepare phase, the SQL-layer acquired
only SNW metadata lock, thus allowing concurrent reads during
this phase (as per WL specification). At the same time InnoDB
assumed that during the prepare phase the table is used only by the
connection executing ALTER TABLE. This mismatch led to assertion
failure in InnoDB.

This fix addresses the problem by ensuring that for in-place
ALTER TABLE variants which require strong lock during the prepare
phase, the SQL-layer acquires an exclusive (X) metadata lock. To
support this change we had to add new variant of locking requirements
for in-place ALTER TABLE - HA_ALTER_INPLACE_SHARED_LOCK_AFTER_PREPARE,
to be used by InnoDB instead of HA_ALTER_INPLACE_SHARED_LOCK
(since InnoDB requires X lock during prepare phase and the
latter acquires only SNW lock for both prepare and main phase).

Commit on behalf of Dmitry Lenev.