[TTS][Magpietts] Added local transformer CFG distillation

[artem-gorodetskii]

Summary

This PR extends online classifier-free guidance (CFG) distillation for MagpieTTS with optional local transformer (LT) distillation.

What’s included

• Added optional local transformer distillation to OnlineCFGDistillation.
• Extended teacher rollout generation to optionally produce local transformer codes and logits.
• Extended student teacher-forced decoding to optionally compute local transformer logits for distillation.
• Added support for local transformer loss computation using the same distillation objectives: KL-divergence, cross-entropy and normalized RMSE.
• Added configuration options to control local transformer distillation: enable/disable LT distillation, LT loss weight, LT distillation start step, LT loss ramp length.

Training flow

• The teacher model is loaded from checkpoint and frozen.
• The teacher performs autoregressive CFG rollout generation.
• When enabled, the teacher also generates local transformer targets and logits aligned with the rollout.
• The generated rollout is fed into the student in teacher-forced mode.
• The student computes decoder logits and, optionally, local transformer logits.
• The total training loss is computed from: the main decoder distillation loss, optional MoE auxiliary loss, optional local transformer distillation loss mixed into the final loss according to the configured LT schedule.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[blisc]

Looks good from my end.