Made message sizes consistent, updated BW calculations

[markstock]

Communications benchmarks typically give the user control over message sizes, to better illuminate crossover points in software and hardware performance. Several of the collective benchmarks were using message sizes of numproccountsizeof(float) bytes. For consistency with other tools and benchmarkers' expectations, I changed those to use message sizes of count*sizeof(float) bytes.

Some bandwidth calculations were also incorrect (and I made some incorrect by the above change), so those are adjusted. For example, during an alltoall each rank sends (numprocs-1) messages, but during an allreduce there are 2*(numprocs-1) total messages sent.

Finally, I updated the Frontier Makefile to use the rccl and mpich_gtl libraries to support GPU-aware MPI.

With these changes, CommBench roughly recreates the performance of https://github.com/ROCm/rccl-tests.