Replace hardcoded PROVIDER_CAPABILITIES with database-driven capabilities

[lbedner]

Summary

The hardcoded PROVIDER_CAPABILITIES dict in app/services/ai/models/__init__.py is redundant now that we have the LLM catalog database tables.

Current State

Hardcoded dict with:

• supports_streaming
• supports_function_calling
• supports_vision
• free_tier_available

PROVIDER_CAPABILITIES = {
    AIProvider.OPENAI: ProviderCapabilities(
        provider=AIProvider.OPENAI,
        supports_streaming=True,
        supports_function_calling=True,
        supports_vision=True,
        free_tier_available=False,
    ),
    # ... etc
}

Proposed Solution

1. Add capability fields to LLMVendor table (or create LLMVendorCapability table)
2. Populate capabilities via the LLM sync ETL from OpenRouter/LiteLLM metadata
3. Replace get_provider_capabilities() to query database instead of hardcoded dict
4. Update ai providers CLI command to use database-driven capabilities
5. Remove PROVIDER_CAPABILITIES dict and ProviderCapabilities model

Benefits

• Single source of truth (database)
• Capabilities auto-update with LLM sync
• No need to manually maintain hardcoded values
• Extensible for new providers without code changes

Files to Modify

• app/services/ai/models/__init__.py.j2 / .jinja
• app/services/ai/models/llm/llm_vendor.py
• app/services/ai/etl/llm_sync_service.py
• app/cli/ai.py.j2 / .jinja