feat: Model attribution dashboard — track primary vs fallback model per turn

[vivekchand]

Closes #300

What

Adds model attribution visibility to the ClawMetry dashboard — answers the question "which model actually responded to my agent's requests?" including silent fallbacks.

How

• New /api/model-attribution endpoint: scans session JSONL files to extract per-model stats (sessions, tokens, cost) and detect primary vs fallback usage based on the configured model in openclaw.json
• Fallback detection: any session whose observed model differs from the configured primary model is counted as a fallback
• Alert when fallback rate exceeds 10%
• Model Attribution panel added to the Tokens tab in both dashboard themes:
  • 🎯 Primary Model card (sessions + tokens + cost)
  • 🔀 Fallback Rate card (red highlight when >10%, amber when >5%)
  • 💸 Fallback Cost card (total tokens + cost burned on fallback models)
  • Sortable model table — all models sorted by token usage with primary/fallback/unknown badges
• loadModelAttribution() JS wired into loadUsage() — loads automatically on every Tokens tab open
• 5 new tests in TestModelAttribution: structure, types, range validation, session count consistency, model fields

Test

python3 -m py_compile dashboard.py  # passes
pytest tests/test_api.py::TestModelAttribution  # 5 tests