Improve the logic to record intermediate results for multi-casting

See the discussion here: https://github.com/citusdata/citus/pull/3410#discussion_r369527510

Basically, we're using `list_append_unique_int()` for `#shardPlacements * # workerNodes` for each intermediate result in the query.

Especially with #3410, the optimization we had where we had short-circuits when we reach all nodes may likely to be hit less times, since we added `writeToLocalFile` check.

Some ideas we've had to use HTAB or bitmap, and run benchmarks.