[FEA] Add experimental multi-threaded BypassMergeSortShuffleWriter

[abellina]

This is a part of: #5039

This task is to add an experimental multi-threaded writer to handle compression and the IO required for writing shuffle blocks on the CPU. We have noticed that writing these blocks to temporary (map_id, reducer_id) files is CPU-bound in the lz4 case, and so adding more threads is beneficial for shuffle-intensive queries.