feat(supertonic): pick a latent-length bucket per piece

[ivan-digital]

Summary

Follow-up to #141. The published LiteRT graphs are fixed at L=64 latent frames (~4.5 s), so a sentence the window cannot hold is split, and the second piece — a separate generation — restarts with sentence-initial prosody. That residual roughness cannot be fixed on the host; it needs a longer window.

soniqo/Supertonic-3-LiteRT now ships an L=128 bucket (vector_estimator_L128.tflite + vocoder_L128.tflite, ≈9 s; exported by speech-models' export_litert.py --latent-frames 64 128). This PR teaches LiteRTSupertonicTts to use it.

What changed

• Latent buckets. *_L{N}.tflite siblings of the base graphs are discovered at construction (latent_buckets()) and loaded on first use, so memory only grows when a long piece needs one. Each piece runs on the smallest bucket whose window holds its predicted duration (choose_latent_bucket(), pure + unit-tested); the planner's window is the largest bucket. A bucket that fails to load is skipped with a log line and the piece falls back to a smaller window.
• Inputs bound by tensor name. The converter permutes a graph's input slots away from the signature order, and differently per toolchain version: the published base graphs bind [text_mask, text_ids, style_dp], a current litert-torch export binds [text_ids, style_dp, text_mask]. Slots are now resolved from serving_default_args_N at load for all three multi-input graphs, with the published order as fallback. Without this, any re-export of the bundle would have broken at run time ("Failed to register input tensor buffer").
• Docs: LiteRTSupertonicTts section in docs/models.md.

Test plan

• choose_latent_bucket unit test in tests/test_supertonic_tokenizer.cpp; LiteRT suite 39/39.
• Base bundle only (no _L128 files): output byte-identical to fix(supertonic): stop stranding sentence tails in their own chunk #141 — both before and after switching to name-based binding.
• With the L=128 pair, real bundle, seed 1234:
  • the 99-codepoint English and 101-codepoint German sentences that fix(supertonic): stop stranding sentence tails in their own chunk #141 had to split now synthesize as one piece (L=90/128, L=97/128) — no seam, no restart;
  • the French paragraph from [Supertonic-3 / LiteRT] Greedy text chunking creates awkward intra-sentence pauses and isolates trailing single words #140: 0 forced splits (was 6 before fix(supertonic): stop stranding sentence tails in their own chunk #141, 1 after it), silences only at the 11 sentence boundaries; the long lead-in runs at L=112/128.
• Parity of the new graphs vs ONNX Runtime (random inputs): vector_estimator_L128 max|Δ| 1.7e-2 (relative 2.7e-3; the published L=64 graph measures 5.4e-3 / 9.2e-4 under the same test), vocoder_L128 2.7e-4.
• Listening check (en): the one-pass sentence was judged clean.

Depends on the L=128 files being published to soniqo/Supertonic-3-LiteRT; without them behaviour is unchanged.