feat(chat): model override per conversation via inline selector

[esafwan]

What

The inline agent/model selector in the chat input now switches the model/provider for the current agent within the same conversation, instead of jumping to a different agent.

How

• Added backend model-override resolution in AgentManager / run_agent_sync / run_agent_stream: when a model override is supplied, the effective provider and model_name are resolved from the AI Model doc.
• Added set_conversation_model_override whitelisted API to persist (or clear) the override on the Agent Conversation.
• Threaded model_override through new_conversation, send_message_to_conversation, audio transcription, file upload/prepare, and the streaming renderer.
• Frontend loads AI Models grouped by provider and forwards the selected override to every send/upload path.
• The selector pill now reflects the selected override even when the popover is closed.
• Blocks modality-mismatched switches (e.g. image/audio history → non-vision model) and shows a cache-reset notice.
• Provider-level filter chips in the selector let users narrow models by provider.
• Modality badges shown on each model item.
• Key-gating: only providers with a configured API key (or local LLMs) are offered; models from unauthenticated providers are hidden to prevent cross-provider auth errors.
• Streaming renderer: parse JSON body once, resolve provider from the override model for cross-provider switches, and include the normalized model in error sanitization so users see specific messages (e.g. model unavailable).

Verification

• Bench migrated, cache cleared, and frontend rebuilt on the isolated chat-model-switching worktree bench at http://localhost:8002.
• Backend resolution test confirms an OpenAI-default agent resolves to Google when a Gemini model override is supplied.
• Selector opens, provider chips filter the list, modality badges render, and only configured providers appear.

Notes

• If a selected model is itself unavailable (e.g. Gemini 2.5 Pro returns 404 for new users), the UI now surfaces a model-specific unavailable message instead of a generic provider error.