refactor: move SNAC warmup into orchestrator.warmup()

fix: warmup SNAC decoder at startup instead of first request

feat: add KV cache FP8 option and torch.compile for SNAC decoder

feat: add VLLM_ATTENTION_BACKEND option (default auto, supports FLASHINFER)

cleanup: consistent logging, remove raw print() calls

fix: enable uvicorn access logs in supervisord

cleanup: remove stale code, fix docstrings, wire up LOG_LEVEL

fix: add missing buffer_ms param to astream signature

fix: detect workers based on SNAC device and CPU cores, fix total_mem typo

feat: expose chunk_size and buffer_ms as optional API parameters

refactor: rename TTS_DEVICE to SNAC_DEVICE, fix streaming for long text, update docs

chore: remove timing decorator and debug endpoints

fix: revert broken SNAC batch decode, add configurable window size

feat: consolidate to single /v1/audio/speech endpoint, add perf optimizations

chore: add TOKENIZER_MODEL to .env.example