Use Pixeltable query routes directly

[pierrebrunelle]

Pixelbot carried two implementations of the same memory and persona reads: declarative @pxt.query functions plus a plain-Python queries.py layer used by legacy FastAPI endpoints. The duplicate layer could drift, and the frontend-facing query routes required a user_id parameter that the frontend did not send.

This change makes the app entry a clearer Pixeltable example:

• declares endpoint-facing queries beside their FastAPIRouter routes in pixelbot/app.py
• serves canonical GET/POST resources at /api/memory and /api/personas, with semantic search at /api/memory/search
• keeps custom FastAPI handlers for validation-heavy writes and user-scoped deletes
• removes queries.py, the versioned v2 reads, the duplicate manual-write alias, and the backend-only download endpoint
• generates memory exports in the client from the canonical query result
• removes ten unused frontend API wrappers and their orphaned types
• reads notification results directly from Pixeltable, scopes both log sources to the local user, caps the request at 100 rows, and removes a duplicate tool-name list
• updates developer guidance, tests, and the bundled SPA

Validation:

• Ruff format/check and mypy pass
• pytest: 13 passed
• frontend lint and 5 contract tests pass
• frontend production build passes; initial chunk remains 271.64 kB
• wheel and sdist build; lock check passes
• pxt schema check passes
• a catalog created from merged main remains fully in agreement: 19/19 models, zero schema operations
• a real pxt service update served health, canonical memory/persona reads, and the notification log; valid responses returned 200, removed v2 paths returned 404, and notification limits above 100 returned 422

No provider or hosted calls were made.