[Attention] Use FA4 for MLA prefill (#34732)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>

benchmarks/attention_benchmarks/common.py

@@ -77,6 +77,7 @@ class MockKVBProj:
         self.qk_nope_head_dim = qk_nope_head_dim
         self.v_head_dim = v_head_dim
         self.out_dim = qk_nope_head_dim + v_head_dim
+        self.weight = torch.empty(0, dtype=torch.bfloat16)
 
     def __call__(self, x: torch.Tensor) -> tuple[torch.Tensor]:
         """
@@ -213,6 +214,7 @@ class BenchmarkConfig:
     use_cuda_graphs: bool = False
 
     # MLA-specific
+    prefill_backend: str | None = None
     kv_lora_rank: int | None = None
     qk_nope_head_dim: int | None = None
     qk_rope_head_dim: int | None = None