Add Qwen3-TTS voice design, consented cloning, and optional fine-tuning

[atomantic]

Part of #5377. Blocked by #5380.

Goal

Add one thoroughly integrated richer local voice backend for original voice design, consented rapid cloning, instruction-controlled synthesis, streaming, and later optional fine-tuning.

Proposed approach

• Add Qwen3-TTS in an isolated, health-probed Python runtime behind the shared voice-engine capability contract.
• Make model downloads, voice design, cloning, and fine-tuning explicit user actions with visible provider/model/resource requirements.
• Add Voice lab flows for same-script design previews, consent and rights confirmation for uploaded/recorded material, source transcription, comparison, promotion, and rollback.
• Support streaming and separately approved studio and interactive routes; benchmark character similarity and first-audio latency before enabling interactive use.
• Reuse the existing training UX pattern for dataset readiness, progress, samples, checkpoints, cancellation, and explicit checkpoint promotion.
• Keep raw recordings, performer data, and artifacts machine-local.

Acceptance criteria

• Voice design and consented instant cloning create candidate profiles without changing the approved character voice.
• Promotion is explicit and records engine/model revision, seed/instructions, source hashes, rights confirmation, and benchmark revision.
• Interactive routing is enabled only after a host-specific latency/readiness benchmark passes.
• Fine-tuning is optional, cancellable, checkpointed, and never assumes the last checkpoint is best.
• Unsupported controls are capability-gated rather than silently approximated.
• No model download, training, or batch generation runs at boot or without direct user consent.

Out of scope

• Adding multiple competing new TTS backends in the first slice.
• Cloning a voice without documented consent or license posture.