Problem: Running insert, alter and drop schema concurrently could
trigger a deadlock.

Root cause was that it was possible to block waiting for a TABLE_SHARE
being opened by another thread while already holding MDL on the
schema. The other thread would, as part of opening the share, attempt
to also acquire MDL on the schema. This could result in a pseudo
deadlock, as the operation would proceed only after the thread
requesting the MDL had timed out.

Solution: When the share is not found in the tdc and there is no prior
MDL on schema, get_table_share() must temporarily release LOCK_open,
acquire MDL on schema, and retry the tdc lookup. This way there is no
need to acquire schema MDL while share->open_in_progress is true. Note
that the MDL now has transaction duration rather than a (shorter) explicit
duration. This should not harm concurrency, as exclusive access to the
schema was always prevented by MDL on a table in it, and concurrent
access is still possible.