fix: handle window_size passed as list

mha.py

@@ -514,6 +514,10 @@ class MHA(nn.Module):
             alibi_slopes = torch.tensor(get_alibi_slopes(num_heads), device=device)
         else:
             alibi_slopes = None
+
+        if isinstance(window_size, list):
+            window_size = tuple(window_size)
+
         if window_size != (-1, -1):
             assert use_flash_attn, "Local (sliding window) attention code path requires flash_attn"