Support richer prompt voice agent configuration

Prompt voice support in PR #9364 intentionally keeps authoring minimal: init scaffolds defaults, and users can currently customize model/modelType, voice, instructions, and store in YAML.\n\nFollow up with richer voice agent configuration support. The final file placement should align with the prompt-agent planning work: simple service-level settings may stay in azure.yaml, while richer agent-specific config may move to agent.yaml / agent definition.\n\nPotential configuration areas:\n- Audio input transcription model, e.g. azure-speech / azure-mrs / mai-transcribe-*\n- Turn detection settings: server_vad threshold, prefix padding, silence duration\n- Audio input/output format settings, including sample rate if service supports it\n- Output voice settings beyond the current simple voice field\n- Session-start greeting / initial response behavior\n- Avatar / video avatar settings\n- Tools and tool resources for voice agents\n- Handoff graph / technical-support or escalation flow\n- Store/logging options beyond the current boolean\n- BYOM validation or optional discovery of compatible deployments\n\nNon-goals for the first follow-up:\n- Do not block the current prompt-voice init/deploy PR.\n- Do not expose a large CLI flag surface before public-preview UX is finalized.\n- Prefer source-controlled YAML configuration over one-off flags.

Read PR #9364 and the prompt-agent planning work first, then inspect how current prompt voice settings are represented in azure.yaml. Determine the supported scope and file placement for richer configuration, and verify that the resulting source-controlled YAML covers the agreed voice-agent settings without expanding the CLI flag surface.

Written by the indexing model from the issue text.