BREAKING CHANGE: All sgl_router_* metrics have been removed.

This commit completes Phase 3 of the metrics migration plan, removing the
legacy metrics system in favor of the new layered SmgMetrics architecture.

## What Changed

Removed the entire RouterMetrics struct (~700 lines) and all sgl_router_*
prefixed metrics from 16 files across the codebase. The new smg_* metrics
with the layered architecture remain as the sole metrics system.

## Why

The legacy sgl_router_* metrics had several issues:
- Inconsistent labeling across different metric types
- No clear separation of concerns between layers
- High cardinality labels in some metrics
- Difficult to correlate metrics across the request lifecycle

The new SmgMetrics architecture provides:
- 6 clear layers: HTTP → Router → Worker → Discovery → MCP → Database
- Consistent labeling with predefined constants (smg_labels module)
- Low cardinality where possible with bounded label values
- Better observability for gRPC streaming (TTFT, TPOT, token counts)
- Support for Responses API (MCP tool calls, database operations)

## Migration Required

Users must update their dashboards and alerts to use the new smg_* metrics.
See .claude/metrics-architecture.md for the full metrics reference and
example PromQL queries.

Old metrics (removed):
  sgl_router_requests_total, sgl_router_request_duration_seconds,
  sgl_router_worker_health, sgl_router_processed_requests_total, etc.

New metrics (use these):
  smg_router_requests_total, smg_router_request_duration_seconds,
  smg_worker_pool_size, smg_worker_requests_active, etc.