Expand supported attention head sizes (#752)

There's no reason for current attention head sizes restrictions - we
theoretically can support any size with current implementations. This
patch fixes that.

vllm/attention/ops/hpu_paged_attn.py

@@ -28,7 +28,7 @@ class HPUPagedAttention:
 
     @staticmethod
     def get_supported_head_sizes() -> List[int]:
-        return [64, 80, 96, 112, 128, 256]
+        return list(range(1, 257))
 
     @staticmethod
     def get_kv_cache_shape(