[Attention] Refactor FA block_size limitations to hybrid models only (#29084)

Signed-off-by: NickLucche <nlucches@redhat.com>

vllm/v1/attention/backends/rocm_aiter_fa.py

@@ -447,7 +447,10 @@ class AiterFlashAttentionMetadataBuilder(
 class AiterFlashAttentionBackend(AttentionBackend):
     accept_output_buffer: bool = True
     supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16]
-    supported_kernel_block_sizes: ClassVar[list[int | MultipleOf]] = [MultipleOf(16)]
+
+    @staticmethod
+    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+        return [MultipleOf(16)]
 
     @classmethod
     def get_supported_head_sizes(cls) -> list[int]: