[TTS][Magpietts] Added CFG distillation

[artem-gorodetskii]

Summary

This PR adds support for online classifier-free guidance (CFG) distillation for MagpieTTS.

What’s included

• Added a new OnlineCFGDistillation model that extends MagpieTTSModel for online CFG distillation training.
• Added MagpieTTS distillation loss implementations, including KL-divergence, cross-entropy, and normalized RMSE losses.
• Added a new training mode, online_cfg_distillation_train, to examples/tts/magpietts.py.

Training flow

• Teacher model is loaded from checkpoint and frozen.
• Teacher performs autoregressive CFG rollout generation.
• Generated rollout is fed into the student in teacher-forced mode.
• The student is optimized using a weighted combination of KL-divergence, cross-entropy, normalized RMSE, and optional MoE loss.

[github-actions]

[🤖]: Hi @artem-gorodetskii 👋,

We wanted to let you know that a CICD pipeline for this PR just finished successfully.

So it might be time to merge this PR or get some approvals.

[github-actions]

[🤖]: Hi @artem-gorodetskii 👋,

We wanted to let you know that a CICD pipeline for this PR just finished successfully.

So it might be time to merge this PR or get some approvals.

[github-actions]

[🤖]: Hi @artem-gorodetskii 👋,

We wanted to let you know that a CICD pipeline for this PR just finished successfully.

So it might be time to merge this PR or get some approvals.