fix(qwen3_tts): reject unsupported voice on Base models instead of silently ignoring it

Base checkpoints ship an empty spk_id table (voice conditioning only
works via ref_audio + ref_text cloning), but generate() never
validated `voice` on that path, so e.g. --voice Chelsie silently
produced fully unconditioned, different-sounding audio on every call.
Mirrors the existing validation already done in generate_custom_voice().

Also fixes README/docs/CLI-help examples that told users to pass
--voice Chelsie/Ethan to Base models — those aren't even real Qwen3-TTS
speaker names (that's a Qwen-Omni voice, not this model family's).

Fixes #892

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>