[Feature]: Support tp=4 when number of heads = 32 for Deepseek MLA aiter kernel

A community user is able to fit Deepseek R1 model on 4 GPUs on MI325x.

We were able to recreate the issue on MI300x.

When aiter MLA kernel is enabled, it would throw exception

Example error log in vLLM V0:

[rank0]:   File "/app/vllm/vllm/attention/ops/rocm_aiter_mla.py", line 67, in mla_decode_fwd_impl
[rank0]:     mla_decode_fwd(q,
[rank0]:   File "/app/aiter/aiter/mla.py", line 116, in mla_decode_fwd
[rank0]:     num_kv_splits, mgc = get_meta_param(num_kv_splits, device, bs, nhead)
[rank0]:                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/app/aiter/aiter/mla.py", line 89, in get_meta_param
[rank0]:     assert nhead in get_mgc, f"{nhead=} not supported"
[rank0]:            ^^^^^^^^^^^^^^^^
[rank0]: AssertionError: nhead=32 not supported

Example error on vLLM V1:

AssertionError: Aiter MLA only supports 16 or 128 number of heads.
Provided 32 number of heads.
Try adjusting tensor_parallel_size value.

Checked the aiter source code, assertion in source code :
https://github.com/ROCm/aiter/blob/main/aiter/mla.py#L159

Note:
This issue is created to follow up on vLLM issue: https://github.com/vllm-project/vllm/issues/19332.

Operating System

No response

GPU

MI300 series

ROCm Component

AITER MLA kernel

Start at aiter/mla.py line 159, where the assertion rejects 32 heads, and trace the metadata path through get_meta_param and mla_decode_fwd. Reproduce the DeepSeek MLA case with tensor parallelism 4 on the documented MI300-series setup; done means the 32-head configuration no longer raises the unsupported-head assertion.