Update bandwidth and latency calculations, add multi work group support

[avinashkethineedi]

• Refined bandwidth and latency calculations for improved accuracy
• Added multi work group support for functional tests

[Yiltan]

Looks good, mostly very minor comment or question about my understanding

[Yiltan]

1. I've seen this pattern a few times in this PR, is there a clean way to abstract this? There may not be
2. I was under the impression that #teams == #work group, if so
   2.a. This could be quite slow for RO as we will be calling MPI_Comm_dup 39 times but we may only need 4 teams for a specific functional test (i.e., mpirun -np 4 ./rocshmem_example_driver -w 4 ... )
   2.b If these assumptions are true, is it possible to only create the number of teams that the test requires?
3. rocshmem_team_split_strided should return a team that is in device memory, is it possible that we create the team_alltoall_world_dup array in device memory so that we don't have to issue that memcopy?

[avinashkethineedi]

• For collective operations, the number of teams should match the number of work groups participating in the operation. This may slow down the init phase for the RO backend, as it requires creating multiple MPI windows.
• Updated the logic so that the number of teams created will now be determined by the ROCSHMEM_MAX_NUM_TEAMS environment variable if set; otherwise, it defaults to 39, since the default number of teams is 40.
• Removed memcpy and moved the team_alltoall_world_dup array to device memory.

[Yiltan]

In line 88, we could allocate as char as we consistently cast to it

[avinashkethineedi]

The r_buf and s_buf buffers are declared as int because they store contiguous integer values used for result verification after the rocSHMEM RMA APIs transfer data to r_buf. However, within the kernels, the getmem/putmem APIs handle data movement and require the pointer to be cast to char*. This is necessary because these APIs interpret the size parameter as the number of bytes to transfer; without casting, the pointer would be treated as an int*, leading to unintended integer-based pointer arithmetic.