feat: standalone STT models

[Lancetnik]

Is your feature request related to a problem?

Speech-to-text only exists inside a realtime Live session: the provider config turns on
input transcription as part of a full-duplex LiveAgent run. There is no transcription
capability independent of it. So dictating a message into the composer requires standing
up a whole realtime voice session on whichever provider the Active Live model uses, the
transcriber cannot be chosen separately from the voice, and an audio file the user
attaches cannot be transcribed at all.

Describe the solution you'd like

Make STT its own configurable model kind alongside Text and Live — provider plus model,
stored in the shared install-wide store with the same Active/per-profile override
semantics — used for dictation into the composer and for transcribing audio, whether or
not a Live model is configured or Active. Providers register the way voice providers
already do, so adding one is a registration rather than a branch.

Additional context

• src/assistant/voice_providers.py — the provider registry to extend; input transcription today
• src/assistant/voice.py, src/assistant/live_configs.py, src/assistant/llm_configs.py
• web/src/components/settings/VoiceSection.svelte, web/src/components/VoicePicker.svelte
• ADR 0004 (shared LLM config store) for the Active / per-profile override semantics

[fasuizu-br]

Affiliation: I am writing on behalf of BRAINIALL, which operates a metered hosted batch speech-to-text route returning a transcript and available word timings. I will not open a PR, request credentials, process private or real-person audio, or claim AG2 provider compatibility. Since #74 separates STT from realtime Live sessions and defines a provider registry, would one maintainer-selected rights-cleared generated WAV of at most 10 seconds be useful as one external standalone-STT contract observation, with input hash and duration, transcript, available timings, response MIME, elapsed time, observed charge, and explicit no-speech or over-limit outcomes posted here? This would not claim model-store, Active-profile, composer, attachment, event-stream, UI, cancellation, bundle, or AG2 Assistant integration.