Enhance: Stereo and local Model Selection

[sp1np0rtal]

Changes

This PR adds several quality-of-life and usability improvements to the MusicGen node:

• Added support for loading MusicGen models from the local models/musicgen directory instead of requiring downloads from Hugging Face.
• Added a model selection dropdown that automatically lists available local MusicGen models.
• Added stereo output support with a selectable Mono/Stereo option (previously only mono output was available).
• Added explicit CUDA device selection/help alongside the existing Auto/CPU options.
• Updated model caching/reload logic so changing the selected model correctly reloads the requested model rather than continuing to use an already loaded model instance.

Notes

During testing I observed that model selection changes can occur while a valid model object is already loaded. For that reason the model reload check remains necessary to detect model swaps and reload the correct checkpoint when the selected model changes.

Tested with local musicgen-medium and musicgen-small checkpoints and verified that changing the model selection triggers loading of the newly selected model.

New features:

• Local MusicGen model dropdown
• Mono/Stereo output selection
• CUDA device selection

Tested with:

• musicgen-small
• musicgen-medium

Existing workflows continue to function as before. If no local model is selected, behavior remains unchanged.