feat: add AI model selection dropdown with per-model completion params

[cjot]

Summary

Adds a model selector dropdown to the assistant onboarding UI so users can choose which WebLLM model to download before clicking "Download AI model".

Changes

Model configuration (web/src/ai/webgpu-provider.ts)

• New ModelOption interface with id, label, description, vramMb, and optional completionParams
• AVAILABLE_MODELS array with two entries:
  • Qwen2.5 0.5B (default, ~945 MB VRAM) — the existing model, no behavioral change
  • Qwen3 1.7B (~2 GB VRAM) — includes completionParams: { extra_body: { enable_thinking: false } }
• WebGPUProvider constructor now accepts ModelOption instead of a raw string
• Passes completionParams through the AiWorkerRequest

Web Worker (web/src/ai/ai.worker.ts)

• Stores completionParams on init, merges into every completion request via spread
• Zero model-specific conditionals — adding a model with new params is one array entry

UI (web/src/ai/ChatPanel.tsx, web/src/App.tsx)

• <select> dropdown shown in the onboarding section when status === 'idle'
• selectedModel state in App.tsx, passed down as props
• handleAiDownload looks up the matching ModelOption and passes it to new WebGpuProvider(config)

Styles (web/src/style.css)

• .chat-model-select styled to match the existing button aesthetic

Testing

• npm run check (tsc) — clean
• npm test (vitest) — 353 passed, 4 pre-existing failures in App.test.tsx (missing assumptions in mock data, unrelated to these changes)
• ruff format / ruff check — clean