Resync all tests with test code from NCCL 2.4

Major rework to merge most of the changes from the NCCL internal
tests into the public ones

Data integrity checking is now much more performant at scale.
Test latency units are now in usec
Startup times at scale are improved

Added a precision for AllGather and ReduceScatter sizes since NCCL us…

…es the size per rank.