Code review for commit 090010f2.

Fix cases where an elog(ERROR) partway through a function would leave the
persistent data structures in a corrupt state.  pgstat_report_stat got this
wrong by invalidating PgStat_TableEntry structs before removing hashtable
entries pointing to them, and get_tabstat_entry got it wrong by ignoring
the possibility of palloc failure after it had already created a hashtable
entry.

Also, avoid leaking a memory context per transaction, which the previous
code did through misunderstanding hash_create's API.  We do not need to
create a context to hold the hash table; hash_create will do that.
(The leak wasn't that large, amounting to only a memory context header
per iteration, but it's still surprising that nobody noticed it yet.)