[ROCm][gfx11] Restore TRITON_ATTN priority on gfx11

[roberteg16]

On RDNA (gfx11/gfx12), the attention backend selector was picking ROCM_ATTN instead of TRITON_ATTN for the auto-selected (no --attention-backend) case. ROCM_ATTN routes decode through chunked_prefill_paged_decode → kernel_paged_attention_2d, which is significantly slower on RDNA than TRITON_ATTN's kernel_unified_attention. On gfx1151 (Strix Halo) this is a ~7× per-decode regression.

This PR:

1. Removes the unconditional ROCM_ATTN prepend in _get_backend_priorities (ROCm) so the existing gfx1x block — which intentionally ranks TRITON_ATTN ahead of ROCM_ATTN — is honored again.

Root cause

Wrong upstream merge of commit 95b4d2b. This commit changed preferred order for gfx11.

_get_backend_priorities (in vllm/platforms/rocm.py) was prepending ROCM_ATTN at priority 0 whenever use_kv_connector was false:

backends = []
# ROCM_ATTN uses (2, num_blocks, ...) KV cache layout which is
# incompatible with KV connectors that require blocks-first layout.
if not use_kv_connector:
    backends.append(AttentionBackendEnum.ROCM_ATTN)
...
if on_gfx1x():
    # On RDNA (gfx11/gfx12), TRITON_ATTN is faster than ROCM_ATTN ...
    backends.append(AttentionBackendEnum.TRITON_ATTN)
    backends.append(AttentionBackendEnum.ROCM_ATTN)

Because get_valid_backends assigns priority by list index (lowest wins), the priority-0 prepend overrode the gfx1x preference, so ROCM_ATTN was selected on RDNA. The prepend's not use_kv_connector guard was also redundant: connector compatibility is already enforced in validate_configuration via supports_kv_connector(), which RocmAttentionBackend overrides to False. Removing the prepend (and the now-unused use_kv_connector parameter) is therefore safe for the connector case and restores correct RDNA ordering.

Profiling evidence (gfx1151, Qwen3 W4A16 MoE)

Same op (vllm::unified_attention_with_output), same input shapes; only the selected backend's kernel differs:

Path	Decode kernel	per-instance	total (×1270 decodes)
TRITON_ATTN (expected)	kernel_unified_attention + reduce_segments	~35 µs	~44 ms
ROCM_ATTN (regressed)	kernel_paged_attention_2d	~248 µs	~315 ms

W4A16 GEMM time was unchanged between the two runs; the entire end-to-end delta came from this attention-kernel difference.

Test plan

• On-device gfx1151 benchmark confirming decode attention returns to the kernel_unified_attention path.

[mgehre-amd]

Your CI failure might be fixed by #991