[DOC] Multi-Threaded shuffle documentation is not accurate on the read side

[abellina]

The Multi-Threaded shuffle documentation says:

The multi-threaded shuffle targets the “BypassMergeSortShuffle” shuffle algorithm in Spark, which is the default when spark.shuffle.partitions is 200 or less.

Unfortunately, that isn't true for the read side. The write side of the shuffle is following this, as Spark has different shuffle writer algorithms (bypass merge, and merge sort). The reader side is a single implementation in Spark, so we don't have a "bypass merge" reader or a "merge sort" reader, it's just the reader. Therefore the documentation should state this, and it's currently incorrect.

Note that we have reduced spark.rapids.shuffle.multiThreaded.maxBytesInFlight lately from 2GB to 128MB because of memory constraints #9153, and this is an ideal knob to control the size in bytes that we allow to be in flight in the decompression/decode threads. Another option to disable the MT reader side only entirely would be to set spark.rapids.shuffle.multiThreaded.reader.threads=0. This is another tool if a user is having issues at shuffle read time only.

[kuhushukla]

Fixed.