PyTorch patches native_layer_norm to fix CUDA mixed-dtype silent errors

PyTorch contributor jansel landed a fix (PR #185693) for a subtle inconsistency in the `native_layer_norm` decomposition. The issue: PyTorch's `aten.native_layer_norm` CUDA kernel requires that non-empty weight and bias parameters have the same dtype as the input tensor. However, the Python reference decomposition (used in tracing, fake tensors, and `aot_eager_decomp_partition`) did not enforce this restriction. It would silently promote mixed dtypes, cast the output back, and accept programs that the actual CUDA kernel would reject with an error. This meant a workflow using `aot_eager_decomp_partition` could generate a graph that would fail at runtime on GPU.

The fix adds a CUDA-specific check in the shared decomposition path (not a backend-specific hook) to validate that weight and bias dtypes match the input dtype when the layer norm has non-empty rows. For empty-row inputs (where no kernel launch occurs), the native behavior is preserved. The patch also includes comprehensive tests: `CUDA repros` tests and CPU decomposition tests ensure correctness. Benchmarking on a bf16 input of shape (32, 128) shows no performance hit — median time actually slightly improved from 79.42 µs to 77.98 µs. This PR, approved by ColinPeppler, resolves issue #151478 and ensures decomposition fidelity matches eager CUDA execution.