fix latency, add dynamic model selector

[Jehu Gray (lowelljehu)]

Latency:

• Cap reasoning_effort for gpt-5/o-series models on raw chat-completions call sites (contextual_agent, semantic_translator tile-selection) to avoid unnecessary reasoning-token overhead on simple prompts.
• Attempt reasoning_effort at Agent Service agent-creation time for reasoning models (analyst_agent), with a safe fallback since the installed azure-ai-agents SDK does not support the parameter.

Model selector:

• AnalystAgent now honors the frontend's selected model end-to-end via a new _ensure_model()/_create_agent_capped() mechanism that recreates the Agent Service agent definition on model change while preserving threads (thread_id is not tied to a specific agent_id).
• ClarifierAgent (greeting/identity short-circuit) now accepts a per-call model override instead of using a fixed singleton deployment, and no longer unconditionally sends reasoning_effort to non-reasoning models.
• Fixed a pre-existing bug where ClarifierAgent used AZURE_AI_PROJECT_ENDPOINT (an Agent Service project endpoint) instead of AZURE_OPENAI_ENDPOINT, causing a silently-swallowed 401 on every greeting.
• Added GET /api/models: discovers real Azure OpenAI deployments via the ARM management API (with a 5-minute in-memory cache and a static fallback list), so the dropdown reflects whatever models are actually deployed in a given tenant instead of a hardcoded list.
• ModelSelector.tsx now fetches from /api/models instead of guessing availability from the overall /api/health status.

[Jehu Gray (lowelljehu)]

@microsoft-github-policy-service agree company="Microsoft"