fix(mhc): MHCPreNormFn.backward returns wrong number of gradients

[yurekami]

Summary

MHCPreNormFn.forward (tile_kernels/modeling/mhc/ops/norm_fn.py) takes 5 user inputs (x, fn, norm_eps, fuse_grad_acc, n_splits), but backward returns 6 elements. PyTorch autograd requires one returned gradient per input, so this raises at runtime:

RuntimeError: function MHCPreNormFnBackward returned an incorrect number of gradients (expected 5, got 6)

Fix

Drop the trailing None from both return statements so the arity matches the 5 forward inputs, and update the type annotation to reflect the real return shape (x_grad is None when fuse_grad_acc=True).

-    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, None, None]:
+    ) -> tuple[torch.Tensor | None, torch.Tensor, None, None, None]:
...
-            return None, fn_grad, None, None, None, None
-        return x_grad, fn_grad, None, None, None, None
+            return None, fn_grad, None, None, None
+        return x_grad, fn_grad, None, None, None