add nn.WeightNorm layer

[mm65x]

Proposed changes

#1888

adds nn.WeightNorm, a module wrapper that applies weight normalization to a
parameter of a given module. reparameterizes a weight w into a magnitude
weight_g and direction weight_v such that w = g * v / ||v||, recomputed
on each forward pass.

implemented as a pure nn.Module layer with no C++ or free functions, as
suggested in #1921. the wrapped module's original weight is frozen so only
weight_g and weight_v are trainable, using the same freeze/unfreeze pattern
as BatchNorm's running stats.

works with any module that has a weight parameter (Linear, Conv1d, Conv2d, etc.).

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)