port(BACKEND-CUDA-COMP-FA): the MLA prefill priority list is the pre-bf2b45b5d6 order and takes no MLA dimensions

Row: BACKEND-CUDA-COMP-FA

Found by re-deriving upstream bf2b45b5d6 (vllm#42669) against this tree in
#2632, wave PORTQ-1.

The gap

LookupMlaPrefillPriority is a 1:1 port of upstream's
_get_mla_prefill_backend_priorities, and it returns the pre-commit order.
include/vllm/platforms/cuda_attn_priority.h:144-149:

inline std::vector<std::string> LookupMlaPrefillPriority(int major) {
  if (major == 10) {  // selector.py:66-72 (Blackwell datacenter)
    return {"FLASH_ATTN", "TRTLLM_RAGGED", "FLASHINFER", "TOKENSPEED_MLA"};
  }
  return {"FLASH_ATTN"};
}

Upstream @ bf2b45b5d6 (a forward reference past the pin 5559679229) gives that
function an mla_dimensions argument (mla/prefill/selector.py:47-80) and, on
Blackwell with the GLM dims, returns
[TRTLLM_RAGGED, FLASH_ATTN, FLASHINFER, TOKENSPEED_MLA] instead
(:170-176). The port takes no dimensions and cannot express the reorder.

Reached from src/vllm/platforms/cuda.cpp:163-165
(get_mla_prefill_backend_priority) and exercised by
tests/vllm/v1/attention/test_attn_backend_registry.cpp:98.

What is NOT in scope, and why the rest of the commit is inert

The commit has three parts. Two of them are FA4-gated and unreachable here:
fa_utils.py:134,231-255 (the requires_local_attention demotion and the widened
TMEM exception) and mla/prefill/flash_attn.py:288-307 (dropping the FA4 branch
from supports_mla_dimensions). This tree ships FA2 only, recorded at
include/vllm/v1/attention/backend.h:445-448; get_flash_attn_version and
supports_mla_dimensions have no counterpart here at all.

Honest weight

Mirror fidelity, not a behaviour change on any board present. Only
FLASH_ATTN is implemented here, and this tree's device is sm_121 (major 12),
so the major == 10 branch the commit reorders is not taken on hardware anyone
runs today. A reviewer may reasonably down-rank this. It is filed because the
divergence sits inside a function whose whole purpose is to mirror upstream's
selector, and because a stale priority list is exactly the kind of thing that is
discovered late, on the first Blackwell datacenter board.

Size

~15-25 lines: an MLA-dimensions parameter and the GLM branch in
include/vllm/platforms/cuda_attn_priority.h, threading the dims through the
virtual in src/vllm/platforms/cuda.cpp and
include/vllm/platforms/interface.h:485-489, plus one registry test case.

Start with LookupMlaPrefillPriority in include/vllm/platforms/cuda_attn_priority.h and trace get_mla_prefill_backend_priority through src/vllm/platforms/cuda.cpp and include/vllm/platforms/interface.h. Use tests/vllm/v1/attention/test_attn_backend_registry.cpp as the verification point; done means MLA dimensions reach the selector and the GLM Blackwell case matches the stated priority order.