Replace the final `FP32LayerNorm` and manual shift/scale application with a single `AdaLayerNorm` module in both the `WanTransformer3DModel` and `WanVACETransformer3DModel`.

This change simplifies the forward pass by encapsulating the adaptive normalization logic within the `AdaLayerNorm` layer, removing the need for a separate `scale_shift_table`. The `_no_split_modules` list is also updated to include `norm_out` for compatibility with model parallelism.

…anVACE transformers

Updates the key mapping for the `head.modulation` layer to `norm_out.linear` in the model conversion script.

This correction ensures that weights are loaded correctly for both standard and VACE transformer models.

… in Wan and WanVACE transformers

Replaces the manual implementation of adaptive layer normalization, which used a separate `scale_shift_table` and `nn.LayerNorm`, with the unified `AdaLayerNorm` module.

This change simplifies the forward pass logic in several transformer models by encapsulating the normalization and modulation steps into a single component. It also adds `norm_out` to `_no_split_modules` for model parallelism compatibility.

Corrects the target key for `head.modulation` to `norm_out.linear.weight`.

This ensures the weights are correctly mapped to the weight parameter of the output normalization layer during model conversion for both transformer types.

Adds a default zero-initialized bias tensor for the transformer's output normalization layer if it is missing from the original state dictionary.