LoRA serving: Observability — Prometheus metrics and logging

[zeyuyuyu]

Context

Parent issue: #367 (Serve fine-tuned models to end-users)
Code review on PR #379 identified missing observability as a medium-priority gap.

Scope

Add Prometheus metrics and structured logging for LoRA adapter lifecycle:

Metrics

• lora_adapters_total (gauge): Number of adapters by state (active/loading/offloaded/failed)
• lora_adapter_deploy_duration_seconds (histogram): Time to download + decrypt + deploy
• lora_adapter_operations_total (counter): Operations by type (deploy/offload/restore/fail)
• lora_event_watcher_lag_blocks (gauge): current_block - last_processed_block
• lora_storage_download_bytes_total (counter): Bytes downloaded from 0G Storage
• lora_inference_requests_total (counter): Requests by adapter name

Logging

• Structured fields for adapter operations (taskID, userAddress, adapterName)
• Request tracing through proxy → ownership check → backend

Acceptance Criteria

• Metrics exposed on /metrics endpoint
• Grafana dashboard template for LoRA serving
• Alert rules for failed deployments and high event watcher lag