fix(stt): ship GGML Whisper models — the shipped ONNX exports can never load

[BernardMasika]

What this fixes

STT never worked as shipped: WhisperManager.InitModel fails on every model the install docs deliver. This is bug B3 from the integration test report — the STT stage was dead end-to-end for any consumer.

Root cause

A format mismatch between the models and the engine. Macoron/whisper.unity wraps whisper.cpp, whose native model format is single-file GGML (ggml-*.bin from ggerganov/whisper.cpp). What the setup script downloads is the onnx-community multi-file ONNX export (encoder + decoder + tokenizer + configs) — a format whisper.cpp has no loader for. The samples then point WhisperManager.ModelPath at encoder_model_quantized.onnx, which can never load.

The fix

Swap every STT surface to GGML:

• tools/setup-sauti.sh downloads ggml-tiny.en.bin (75 MB, essential profile) / ggml-small.en.bin (466 MB, --models all) from ggerganov/whisper.cpp, SHA-256-pinned, placed flat under StreamingAssets/VoiceAI/stt/. GGML models are self-contained — tokenizer, config, and both graphs live inside the one .bin, so the five-file-per-variant layout goes away entirely.
• Samples 02 / 05 / 06 resolve a single GGML file via a sttModelFilePreference list and hand it straight to WhisperManager.ModelPath (replacing the subdir + ONNX anchor-file scheme).
• SautiSetupWizard model check now looks for stt/ggml-tiny.en.bin.
• Metadata/docs: ai-models/stt/manifest.json, ai-models/manifest.json, ai-models/stt/README.md, docs/reference/models.md (+ bundle-size matrix), docs/experiments/02|05|06, the architecture asset-flow diagram, SHIP_READINESS.md, and NOTICE all updated. Licensing is unchanged in substance — OpenAI Whisper is MIT; the GGML conversions by Georgi Gerganov are MIT too.

Size trade-off worth knowing: GGML Small is 466 MB vs 239 MB for the ONNX pair (flagship bundle grows ~1.6 → ~1.8 GiB), and GGML Tiny is 75 MB vs 43 MB. The difference buys an STT stage that actually loads; .en variants keep the English-only contract.

Verified (consumer project: Unity 6000.0.70f1, whisper.unity master, desktop CPU)

• ggml-tiny.en.bin loads via WhisperManager.InitModel() in ~1.8 s; the shipped ONNX files fail to load in the same setup.
• End-to-end round trip: fed Kokoro-TTS-generated audio of "Justice is the foundation of the Baraza courtroom." into GetTextAsync → transcribed in ~0.8 s as " Ustice is the foundation of the Barraza courtroom." — Whisper Tiny on synthetic speech, so effectively perfect.
• Compile clean in the consumer project. Note: the EditMode test suite doesn't cover the changed surfaces (samples ship as source, the wizard check is editor-UI) — the live load + transcription above is the meaningful verification for this change.

Note on versioning

I deliberately did not bump package.json — this PR and #1 would collide on it. The changelog entry sits under [Unreleased]; version it as you see fit when releasing (1.3.6 if this lands after #1).