[TorchOnnxToTorch] Fix LayerNormalization scale/bias dtype mismatch with stash_type

[bjw-0]

Summary

When ONNX LayerNormalization has stash_type != input dtype (e.g., stash_type=1 (float32) with float16 inputs), the existing lowering correctly casts x to the stash dtype but passes scale and bias through unchanged. This causes torch.aten.native_layer_norm to receive mixed dtypes (f32 input with f16 scale/bias), which fails verification or produces incorrect results at runtime.

This PR fixes the issue by:

1. Casting scale and bias to the stash dtype when it differs from the input dtype
2. Using the stash dtype for the native_layer_norm output type
3. Casting the result(s) back to the original output dtype

This complements the stash_type support added in #3888 and is the same class of fix as #4474 (which addressed SimplifiedLayerNormalization).

Fixes #3725

Test plan

• Added two FileCheck tests in simple_ops_g_to_p.mlir:
  • test_layer_norm_stash_type_f16: Single-result case — verifies x, scale, and bias are all cast to f32, native_layer_norm runs in f32, and result is cast back to f16
  • test_layer_norm_stash_type_f16_3results: Three-result case — verifies all three outputs (y, mean, invStdDev) are cast back to f16
• Existing test_layer_norm and test_layer_norm_single_result tests (f32 input, no stash_type mismatch) remain unchanged and continue to pass

[bjw-0]

@zjgarvey @Groverkss Would you mind taking a look at this? This is the standard LayerNormalization counterpart to the SimplifiedLayerNormalization stash_type fix in #4474. Thanks!