Add safe local provider scaffolding and benchmark workflows

[mohitkale]

Summary

• add structured provider manifests and capability metadata across local TTS and audio-intelligence backends
• add opt-in Qwen3-ASR and VoxCPM2 scaffolding with pinned revisions, local-path-only loading, and isolated worker boundaries
• correct Qwen device handling and Chatterbox multilingual language routing
• add a durable English, Hindi, and code-switched benchmark harness
• make Docker exposure localhost-only by default and derive UI language choices from provider capabilities
• document safe Mac-native, Docker CPU, Colab, and Kaggle experimentation workflows
• add disabled, license-gated experimental metadata for IndicF5 and Indic Parler-TTS

Why

VoiceForge needs a reproducible local-first experimentation path that does not
install packages or download or execute model weights during application
startup. This change establishes explicit provider contracts, safe runtime
profiles, model-revision metadata, and a shared benchmark workflow before
heavier GPU experimentation.

User impact

The studio now reports provider availability and capabilities more accurately,
keeps optional models opt-in, exposes Docker only on localhost by default, and
provides repeatable benchmark and notebook guidance for English, Hindi, and
code-switched evaluation.

Validation

• pytest tests: 65 passed
• ruff check app tests scripts: passed
• git diff --check origin/main...HEAD: passed
• Docker CPU profile: healthy on 127.0.0.1:8089
• browser-driven local smoke tests: OpenVoice and XTTS both produced WAV output successfully

Limitations

• optional checkpoints remain explicit downloads and are not bundled
• large models remain better suited to Colab/Kaggle or a GPU worker
• VoxCPM2 and Indic experiments remain disabled by default
• Kokoro was not added because no implementation was present in this checkout