Add optimized/tflite: portable CPU (LiteRT / TFLite) release

[Cortexelus]

Portable CPU (LiteRT / TFLite) release of the SA3 pipeline, parallel to optimized/mlx and optimized/tensorRT. Runs anywhere ai_edge_litert runs (macOS / Linux, x86 / ARM, via the XNNPACK delegate).

What's here

• scripts/sa3_tflite.py — CLI matching the sa3_mlx / sa3_trt flag names and modes: text-to-audio, audio-to-audio (--init-audio), inpainting (--inpaint-range), CFG (--cfg + --negative-prompt + --apg), plus --threads and --cfg-batched.
• scripts/weights.py — lazy HuggingFace auto-download (same ensure_local + symlink-from-cache pattern as MLX), pulling the fp32 .tflite models from stabilityai/stable-audio-3-optimized/tflite/*.
• models/defs/tflite_pipeline.py — tokenizer + T5Gemma wrapper + pingpong sampler.
• models/tokenizer.model — bundled SentencePiece (4 MB); the .tflite T5Gemma is encoder-only, so the tokenizer ships in-repo (cf. the TensorRT release bundling tokenizer.json).
• install.sh / bootstrap.sh / sa3 wrapper / install.py — mirror the MLX release (uv-based; portable, no Apple-Silicon gate).

Notes

• Baked-I/O variable-length graphs (conditioner + patch/unpatch in-graph). The canonical DiT is variable-batch, so CFG runs cond+uncond as one batch=2 invoke by default (--cfg-batched; ~7-29% faster on Apple-Silicon AMX) or sequential batch=1 (--no-cfg-batched, like TensorRT) — bit-identical.
• Monotonic-rebuild audio-to-audio schedule, self-contained here. The equivalent change to the MLX/TRT samplers is a separate PR (sa3-a2a-monotonic-schedule).
• SAME-L chunked decode (chunk=64, overlap=8); SAME-S decodes whole.

All fp32 except T5Gemma (fp16). Models are already live on HF. Tested end-to-end (HF download to 20 s generation, healthy audio; audio-to-audio verified).