Add local repo loading and strip default inference weight normalization

[JINZIPING]

Title

Reduce inference GPU memory and add optional local model loading

Description

This PR reduces GPU memory usage by stripping weight_norm parametrizations after model weights are loaded, and adds an optional parameter for loading model files and voices from a local repo-shaped directory instead of downloading from Hugging Face.

What changed:

• strip weight_norm by default in inference mode after loading weights
• keep an escape hatch with for_training=True
• add local_repo_dir to load config.json, model weights, and voices/ from a local directory
• keep Hugging Face download as the default behavior when local_repo_dir is not provided

Why:

• weight_norm is useful for training, but it adds extra inference overhead
• some deployments already have model assets on disk and do not want to depend on download-time resolution

Result on my setup:

• GPU memory reduced by about 28% (1016 MB -> 724 MB in my test)