feat[2]: WER-based layer importance (WLI) pruning + cosine baseline

Goal

Iteration 2 of the SPADE implementation: the pruning stage. Compute per-layer importance by leave-one-out layer removal measured with WER (WLI, paper Eq. 1), plus the cosine-based layer importance (CLI) baseline used for ablation, and remove low-importance layers to produce the pruned student backbone.

Acceptance

• spade/pruning/ provides compute_wli (leave-one-out WER per layer), compute_cli (mean cosine distance between layer input/output), and layer selection/pruning utilities
• Backbone supports temporarily disabling blocks (leave-one-out) and pruning returns a new student model whose retained blocks copy teacher weights
• WER scorer abstraction: built-in token-codec scorer decoding generated codes to text; optional Whisper-based scorer for real audio
• Pruning + selection + scorer have passing pytest tests

Notes

Iteration 2 for: SPADE - Structured Pruning and Adaptive Distillation for Efficient LLM-TTS (arXiv:2509.20802)

Start with the existing repository structure and define the interfaces under spade/pruning/ for WLI, CLI, layer selection, pruning, and scorer implementations. Run the current pytest suite before adding tests for leave-one-out disabling, cosine and WER scoring, retained teacher weights, and token-codec decoding. Done means the pruning and selection utilities work with both scorer modes and all related pytest tests pass.