Optimize group_index_select_or_add_2d_kernel by adding a separate codepath for small embedding dimensions

[aryaman-gupta]

This PR optimizes the performance of the group_index_select_or_add_2d_kernel kernel on tables with small embedding dimensions (i.e., num_cols).

For tables with small embedding dimensions, the code is refactored to process multiple rows within the same warp. Two files are changed:

1. fbgemm_gpu/src/sparse_ops/sparse_ops_gpu.cpp - The calculation of the warp_offsets is changed in the host-side code.
2. fbgemm_gpu/src/sparse_ops/sparse_group_index.cu - The group_index_select_or_add_2d_kernel kernel is modified to process multiple rows within a warp for small embedding dimensions.

Benchmark results:

Benchmark 1:

	Forward	Backward
Baseline	10 ms	14 ms
Optimized	7 ms	10.5 ms

Benchmark 2:

	Forward	Backward
Baseline	3 ms	6 ms
Optimized	2 ms	5 ms