feat: upstream model health monitoring

[christianromeni]

Summary

Proactive upstream model health monitoring with 3 configurable probe levels.

Health Check Levels

Level	Probe	Default
health	GET server root (any HTTP response = healthy)	On, 30s
models	GET /models (2xx = ok)	Off, 60s
functional	POST /chat/completions with 1 token	Off, 5m

Backend

• New internal/health/ package with Checker, 3 probe types
• Copy-on-write for thread-safe concurrent results
• Error sanitization (no internal URLs/IPs leaked)
• Minimum interval enforcement (10s health/models, 60s functional)
• Prometheus gauges: model_health_status + model_health_latency_seconds
• API: GET /api/v1/models/health
• Dashboard stats: models_healthy / models_unhealthy / models_degraded
• 15 health checker tests

Frontend

• Health badges on Models page (green/yellow/red dots)
• Model Health section on Dashboard
• Model Performance table uses health latency + TPS (not misleading avg duration)
• useModelHealth hook with 15s polling

Config

settings:
  health_check:
    health:
      enabled: true
      interval: 30s

Test plan

• go test ./... -race 17/17 passed
• npx tsc --noEmit clean
• npm run lint clean
• npm run test -- --run 305/305 passed
• Manual: dolphin-mistral shows "healthy" with 1ms latency
• Manual: unreachable models show "unhealthy"