[Perf]: Establish baseline benchmarks and validate PRD latency targets

[Vraj1234]

Describe the performance problem

The PRD defines latency SLAs that have never been measured:

• Standard API queries: < 1s (95th percentile)
• Complex spatial queries: < 2s (95th percentile)
• Cloud Run cold start: < 1s

No benchmark suite exists, so there is no way to know whether these targets are currently met or violated.

Steps / workload to reproduce

1. Stand up the staging stack
2. Run a load test (e.g. locust or k6) against representative endpoints:
   • GET /v1/feedstocks/usda/census/crops/{crop}/geoid/{geoid}/parameters/{parameter}
   • GET /v1/feedstocks/analysis/resources/{resource}/geoid/{geoid}/parameters
   • All discovery endpoints
3. Record p50/p95/p99 latency and error rate
4. Test cold-start latency by letting the Cloud Run instance scale to zero, then sending a request

Software version

Current staging HEAD (main branch)