feat[1]: Project scaffolding, LLM-TTS backbone, and synthetic codec data

Goal

Iteration 1 of the SPADE implementation: set up the Python package skeleton and build the foundation the pruning/distillation stages depend on — a minimal GPT-style LLM-TTS backbone and the data layer (text tokenizer + synthetic speech-codec benchmark that is decodable back to text so WER is meaningful).

Acceptance

• pyproject.toml, .gitignore, and spade/ package skeleton exist with minimal dependencies (torch, numpy, pyyaml, tqdm)
• spade/models/llm_tts.py provides a configurable autoregressive LLM-TTS backbone (causal attention, residual blocks, LM heads) returning logits + per-block hidden states + attention maps
• spade/data/ provides a character-level text tokenizer and a synthetic speech-codec dataset: text -> deterministic code-token sequences, decodable back to text
• Model forward/generation and data pipeline have passing pytest unit tests

Notes

Iteration 1 for: SPADE - Structured Pruning and Adaptive Distillation for Efficient LLM-TTS (arXiv:2509.20802)

Start with the acceptance checklist, then review pyproject.toml and the proposed spade/ package layout. Read the LLM-TTS requirements in spade/models/llm_tts.py and the tokenizer and synthetic dataset requirements under spade/data/. Done means the configurable model and decodable data pipeline exist, with passing pytest coverage for forward/generation and data behavior.