feat[3]: Adaptive multi-level distillation (Skew-KL losses, dynamic layer matching, trainer)

Goal

Iteration 3 of the SPADE implementation: the recovery stage. Implement the paper's composite distillation loss (Eq. 2): CE + Skew-KL logit alignment + latent/attention/embedding MSE, with dynamic layer matching so each student layer distills from the last teacher layer before the next retained layer.

Acceptance

• spade/distill/losses.py provides CE, Skew KL (DistiLLM forward/reverse, masked), latent/attention/embedding MSE, and the SPADE composite loss (alpha=0.25 default)
• spade/distill/matching.py provides dynamic teacher-layer targets (last layer before next retained; final layer for the last student layer) and the static ablation
• spade/distill/trainer.py trains the student against a frozen teacher with the composite loss and returns per-component metrics
• Losses, matching, and trainer have passing pytest tests (including teacher frozen / student updated)

Notes

Iteration 3 for: SPADE - Structured Pruning and Adaptive Distillation for Efficient LLM-TTS (arXiv:2509.20802)

Start by reading Eq. 2 in the cited SPADE paper, then inspect the distillation work under spade/distill/. Implement and test losses.py, matching.py, and trainer.py against the listed acceptance criteria, including frozen-teacher and updated-student behavior. Run the pytest suite and confirm per-component metrics and both dynamic and static matching tests pass.