Qvac 17489 io binding for kv cache chaining

[Zbig9000]

What problem does this PR solve?

• Chatterbox's autoregressive LM loop copied every present.* tensor back into its matching past_key_values.* input on every step, a per-step O(past_len × heads × head_dim) copy that grew linearly with sequence length (quadratic total cost for an N-step generation).
• That per-step copy was a measurable share of LM generation RTF, particularly on the English (non-CFG) path where LM dominates total time.

How does it solve it?

• Add setOutputToInputChain / clearChainedInputs / isInputChained to IOnnxInferSession. After each run() the session std::moves output Ort::Values directly into the matching input slots — no copy, no intermediate user-space vector.
• OnnxInferSession::initInputTensors() preserves chained slots across re-init so the moved tensors survive.
• ChatterboxEngine::enableKvCacheChaining() builds the {present.i → past_key_values.i} mapping from session I/O names (inputs [keyValueOffset_..] map 1:1 onto outputs [1..], since outputs[0] is logits). Wired into both generateSpeechTokens (non-CFG) and generateSpeechTokensWithCfg (CFG) paths. writeKvToTensors and cachePastKeyValues now skip chained inputs.
• Add ChatterboxConfig.kvCacheChaining (default true) wired through TTSModel, AddonJs, and index.js so the optimization can be toggled for A/B benchmarking or an emergency disable.

How was it tested?

• Full unit suite green: 214 passing / 0 failing (2 unrelated LavaSR benchmark tests skipped), qvac-lib-inference-tts-unit-test with integration filters excluded.
• New coverage:
  • 5 KvCacheChainingTest cases: English offset=3 mapping, multilingual offset=2 mapping, truncated-outputs edge case, writeKvToTensors skip, cachePastKeyValues skip.
  • 3 OnnxInferSessionMockTest cases for the new mock methods.
• A/B benchmark on device (Linux, 4 cores, q4 models, jfk.wav reference, kvCacheChaining toggled):
  • English (non-CFG): mean RTF 1.245 → 1.102 (−11.5%), totalTime −12.1% over 3 runs.
  • Multilingual ES (CFG): mean RTF 4.145 → 3.987 (−3.8%), totalTime −3.8% over 2 runs (smaller share because the conditional decoder dominates multilingual).
• Listened to audio output with chaining ON on both paths — no artifacts vs the pre-change baseline.
• JS lint (standard) clean; C++ lints clean.

API Changes
New non-breaking kvCacheChaining option on the ONNXTTS constructor (default true, i.e. optimization on).

const model = new ONNXTTS({
  engine: 'chatterbox',
  files: { /* ... */ },
  referenceAudio,
  config: { language: 'en' },
  kvCacheChaining: false // opt-out, e.g. for benchmarking or emergency disable
})

[GustavoA1604]

Need not to add kvCacheChaining as a JS option. Can just keep it to true always. We can compare with rpevious version by running old version of the addon

[GustavoA1604]

/review

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (1/1)
- 1 Team Lead OR Management approval ✅ (1/1)



---
*This comment is automatically updated when reviews change.*