refactor(api): Move benchmark data processing from UI to backend

[anfredette]

Description

The UI currently receives raw benchmark data from the backend and processes it client-side. This processing should be moved to the backend to keep the UI simpler and more maintainable.

Current Behavior

• load_206_models() in app.py:1600-1624 fetches raw benchmark data from GET /api/v1/benchmarks
• Returns the full dataset as a pandas DataFrame
• UI then processes this data for:
  1. Displaying a searchable model catalog in the About tab (app.py:2769-2792)
  2. Showing model count in spinner messages (app.py:4762)

Problems with Current Approach

• UI coupled to data schema - if benchmark CSV columns change, UI code breaks
• Duplicate logic risk - multiple UI components extracting similar info leads to inconsistency
• Harder to test - data transformations in UI are harder to unit test than backend code
• Larger payloads - full benchmark data transferred when UI only needs a subset
• UI complexity - UI handles parsing, filtering, and error cases that belong in backend

Proposed Fix

Move data processing to the backend:

1. Create purpose-specific backend endpoints that return only what the UI needs:

   • Model catalog endpoint (e.g., GET /api/v1/model-catalog) - returns model names and display metadata
   • Model count endpoint or include count in existing responses

2. Update UI to consume these endpoints directly without DataFrame processing

3. Remove load_206_models() and related DataFrame handling from UI

Benefits

1. Simpler UI - just renders what it receives, no data transformation logic
2. Single source of truth - data processing logic lives in one place (backend)
3. Decoupled from schema - backend can adapt to data source changes without UI updates
4. Smaller payloads - transfer only what's needed for each view
5. Easier testing - backend transformations can be unit tested independently

Files Affected

• ui/app.py - Remove load_206_models() and DataFrame processing
• backend/src/api/routes.py - Add new endpoint(s) for processed data