feat: Add inference metrics

[gyliu513]

What does this PR do?

Fixed #5321

Summary

• Add three new OpenTelemetry inference metrics to track LLM serving performance:
  • llama_stack.inference.duration_seconds end-to-end inference latency (streaming and non-streaming)
  • llama_stack.inference.time_to_first_token_seconds time to first content token (streaming only)
  • llama_stack.inference.tokens_per_second output token throughput (completion_tokens / duration)
• All metrics carry model, provider, stream, and status attributes
• Add Grafana dashboard and kind cluster deployment script

Query Examples

# P50 Tokens Per Second
curl -s --get 'http://localhost:9090/api/v1/query' \
  --data-urlencode 'query=histogram_quantile(0.50, sum by (le) (rate(llama_stack_llama_stack_inference_tokens_per_second_bucket[5m])))'

# P95 Inference Duration
curl -s --get 'http://localhost:9090/api/v1/query' \
  --data-urlencode 'query=histogram_quantile(0.95, sum by (le) (rate(llama_stack_llama_stack_inference_duration_seconds_bucket[5m])))'

# P95 Time to First Token
curl -s --get 'http://localhost:9090/api/v1/query' \
  --data-urlencode 'query=histogram_quantile(0.95, sum by (le) (rate(llama_stack_llama_stack_inference_time_to_first_token_seconds_bucket{stream="true"}[5m])))'