Summary

This PR adds a comprehensive how-to guide for tuning the Python online feature server
in production Kubernetes deployments with the Feast Operator.

The guide covers:

• Worker/connection/timeout tuning — Gunicorn worker sizing, connection limits,
  keep-alive, and max-requests recycling with Operator CR and CLI examples
• Registry cache tuning — explains the synchronous refresh latency problem and
  the cache_mode: thread fix for background refresh
• Online store selection — comparison of Redis, DynamoDB, PostgreSQL, Bigtable,
  and Cassandra with latency guidance and store-specific tuning (connection pooling,
  batch parallelism, retry modes, consistent reads)
• Batch size tuning — per-store batch size comparison and profiling guidance
• ODFV optimization — native Python mode vs pandas, write-time transforms
• Horizontal scaling — expected scaling efficiency table and HA summary
  (links to scaling-feast.md for full YAML examples)
• Metrics setup — Prometheus built-in metrics (key metrics table, alert rule
  examples) and OpenTelemetry Operator integration
• Network latency optimization — co-location, VPC endpoints / Private Link
  for AWS/GCP/Azure, DNS caching

The guide is registered in docs/SUMMARY.md under How-to Guides, after
"Feast on Kubernetes". It cross-links to existing reference docs (scaling,
feature server, OpenTelemetry, store-specific pages, ODFVs).