Revise new_group() implementation for 'mpi' backend

[RadhaGulhane13]

Problem :

Currently, mcr-dl's new_group() implementation for mpi uses torch.distributed to create new_group instead of having it's own implementation.
Refer following :

MCR-DL/mcr_dl/mpi.py

Line 74 in 870751f

def new_group(self, ranks):

We would like to have direct implementation instead of using torch.distributed similar to existing implementation of 'nccl'
Refer following :

MCR-DL/mcr_dl/nccl.py

Line 73 in 870751f

def new_group(self, ranks):

Testing :

• Add new benchmarks at https://github.com/OSU-Nowlab/MCR-DL/tree/main/tests to test these changes
• Test changes for 'V23.05 Megatron-LM with mcr-dl support Megatron-LM#1' with backend='mpi' and distributed_engine='torch'/'mcr-dl'