feat(voice/ux): flatten settings into How It Sounds / How It Responds

Discovered during dogfooding that the advanced settings disclosure mixed
two unrelated concerns: acoustic delivery and conversational behaviour.

Remove the top-level "Advanced voice settings" collapsible entirely.
Replace with two flat, always-visible subsections:

  How It Sounds — voice backend, voice picker, acoustic tuning sliders
    (stability, expressiveness, speaking rate, similarity, affective
    dialog, per-backend hints). All controls that affect the audio signal.

  How It Responds — language, proactive toggle, identity, character
    prompt, delivery preset, platform fixtures. All controls that affect
    what the agent chooses to say and in what language.

Language moves from above How It Sounds (where it was awkwardly grouped
with the backend picker) into How It Responds, where it belongs.

Proactive toggle moves from the separate "Session behavior" block into
How It Responds — it is a conversational behaviour control, not an
acoustic one.

VoiceAdvancedControls exports two named components (VoiceRespondsControls,
VoiceSoundsControls); the old VoiceAdvancedControls export is kept as a
deprecated re-export for any external callers.

zh-CN locale: 声音效果 / 对话风格.

Author: heavygee