Joins with a final scan count wait_probe_counter down to zero before the
probes finish, so the last probe never saw 1 in finish_build and the build
side and hash table were kept until the query ended. Count unfinished probes
separately. The hash table entry storage also lived on the build state and was
never cleared, growing across spill rounds; keep it next to the table and
clear it whenever the table is replaced or released.

The set of seen key digests grows with the input and was kept after the
processor finished, through the segment rewrite and commit that follow.

The last partition's blocks are only released when a new partition starts,
so they were kept until the query ended. Without PARTITION BY that is the
whole input.

Closing the channel keeps queued payloads alive while any sender exists, so a
consumer that stopped early held its unread payloads until the query ended.
Also drop buffered payloads once every consumer is gone.

The probe and build processors kept the shared state, with the concatenated
build side and the packed R-tree, until the query ended. Take it when each
processor finishes so it is freed when the last one is done.

…inish

The partition stream shared by partial aggregates reserves each bucket's last
flushed size again after a flush, and finishing only finalized buckets that
still held rows, so the reserved capacity stayed until the query ended. Drop
every bucket builder once the last partial finishes.