Enable Triton MLA for prefill

[cagrikymk]

This PR adds a flag (VLLM_ROCM_USE_AITER_TRITON_MLA) that enables Triton MLA when the flag is turned on.

The corresponding PR in aiter: ROCm/aiter#1203

[ZJLi2013]

looks a few gaps 1) it's not aligned with main branch; 2) when cherry-pick to main branch, still got runtime error for gfx950(mi35x):

 05:04:23 [multiproc_executor.py:585]     aiter_triton_fp8_bmm(x,^M
 05:04:23 [multiproc_executor.py:585]   File "/usr/local/lib/python3.12/dist-packages/aiter-0.1.5.dev196+gb5f0b0a05.d20251016-py3.12.egg/aiter/ops/triton/batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant.py", line 315, in batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant^M
 05:04:23 [multiproc_executor.py:585]     _batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant_kernel[^M
 05:04:23 [multiproc_executor.py:585]   File "/usr/local/lib/python3.12/dist-packages/triton/runtime/jit.py", line 570, in run^M
 05:04:23 [multiproc_executor.py:585]     options = backend.parse_options(kwargs)^M
 05:04:23 [multiproc_executor.py:585]               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^M
 05:04:23 [multiproc_executor.py:585]   File "/usr/local/lib/python3.12/dist-packages/triton/backends/amd/compiler.py", line 124, in parse_options^M
 05:04:23 [multiproc_executor.py:585]     return HIPOptions(**args)^M
 05:04:23 [multiproc_executor.py:585]            ^^^^^^^^^^^^^^^^^^^M
 05:04:23 [multiproc_executor.py:585]   File "<string>", line 24, in __init__^M
 05:04:23 [multiproc_executor.py:585]   File "/usr/local/lib/python3.12/dist-packages/triton/backends/amd/compiler.py", line 74, in __post_init__^M
 05:04:23 [multiproc_executor.py:585]     assert self.kpack == 1, "gfx950 only accepts kpack == 1"^M

[cagrikymk]

@ZJLi2013 That might be about the Triton version, can you try: https://github.com/ROCm/triton/tree/pytorch/rocm7.1_internal_testing

Also, I have this branch for the vllm upstream for testing: https://github.com/ROCm/vllm/tree/cagri/triton_MHA_upstream

I will also run more tests to see if there are any issues.

[dllehr-amd]

Approved! Merging into 355_wip