Memory leak on hashed agg rescan

I noticed a minor leak in the per-query context when ExecReScanAgg()
is called for a hashed aggregate. During rescan, build_hash_table() is
called to create a new empty hash table in the aggcontext. However,
build_hash_table() also constructs the "hash_needed" column list in
the per-query context, so repeated calls of build_hash_table() result
in leaking this memory for the duration of the query.

Attached is a patch that fixes this by only constructing "hash_needed"
if it doesn't already exist. I also bms_free'd the temporary BMS that
is created, although that's pretty harmless.

Neil