feat[5]: Hugging Face backbone adapter (prune, WLI bypass, distillation)

Goal

Iteration 5 of the SPADE implementation: make the framework directly applicable to HF-style residual Transformer backbones (GPT-2 pattern; same structure as CosyVoice 2 / LLaSA stacks). Provide layer-stack detection, state-dict-remapping layer pruning, leave-one-out block bypass for WLI, and a distillation trainer that reuses the existing SPADE losses via HF hidden states/attentions.

Acceptance

• spade/adapters/hf.py detects a model's layer ModuleList (transformer.h / model.layers / ...), prunes layers by state-dict remapping, and supports temporarily bypassing a block (WLI)
• compute_wli accepts a pluggable bypass function so it works for any backbone
• HFDistillTrainer runs the SPADE composite loss on HF-style outputs (hidden_states/attentions/embedding) with dynamic layer matching
• Tests with a locally-constructed tiny GPT-2 (no downloads): pruning correctness, WLI wiring, teacher frozen / student updated

Notes

Iteration 5 for: SPADE - Structured Pruning and Adaptive Distillation for Efficient LLM-TTS (arXiv:2509.20802)

Start with the acceptance criteria and inspect the existing SPADE losses and compute_wli entry point before creating spade/adapters/hf.py. Define tests around a locally constructed tiny GPT-2: verify layer detection and pruning, WLI bypass wiring, and that the teacher stays frozen while the student updates. Done means HFDistillTrainer handles hidden_states, attentions, embedding, and dynamic layer matching.