[DRAFT] Optimize on lightnling Indexer Triton kernels

[leonling-ll]

This is a DRAFT PR which aim show some optimization reference to the lightenling Indexer Triton kernels proposed in #606.

The main changes are in transformer_engine/jax/triton_extensions/indexer.py
The remained files under standalone_indexer are benchmark scripts mirror to same named files in benchmarks, but remove the dependencies of TE and jax for the ease of testing.

The performance before and after this change (shapes following the benchmark scripts):

Kernel	Shape (B, oH, T, S, H, d_i)	Baseline	Optimized	Speedup
_score_reduce_kernel	2, 64, 4096, 4096, 64, 128	49.1 ms / 722 TF	39.22 ms / 904 TF	1.25x
_score_dscores_chunk_kernel	2, 64, 1024, 1024, 64, 128	12.1 ms / 183 TF	9.45 ms / 235 TF	1.28x
_score_topk_kernel*	2, 64, 1024, 1024, 64, 128, k=512	8.2 ms / 271 TF	5.14 ms / 431 TF	1.59x