Document benchmark limitations and fix CI timeout issues

[tomtom215]

Summary

This PR adds comprehensive documentation for known benchmark measurement limitations and resolves Criterion.rs timeout warnings by increasing measurement time budgets for 5 benchmark groups that were marginally exceeding their time limits on CI runners.

Key Changes

Documentation Improvements:

• Added 8 new "Known Measurement Limitations" entries to benches/README.md and auto-generated benchmark results page covering:
  • Data volume/save wide confidence intervals from BTreeSet rebalancing
  • Dispatch routing inverted results (direct handler vs HTTP round-trip)
  • Cold start vs steady state measurement differences
  • Subscribe fan-out O(1) scaling behavior
  • Agent burst sub-linear scaling with Tokio work-stealing
  • Tenant resolver negligible overhead (88–173ns)
  • Pagination context index 2× speedup via BTreeSet filtering
• Updated CHANGELOG.md with performance highlights from the cleanest benchmark run in project history (237 benchmarks, zero panics/errors)
• Added benchmark overview section to book/src/deployment/testing.md documenting all 13 suites and 237 benchmarks
• Updated CI/CD documentation in book/src/deployment/cicd.md describing the benchmarks workflow

Benchmark Timeout Fixes:

• transport/payload_scaling: 8s → 10s (4KB and 16KB payloads needed 8.4–9.5s)
• concurrent/sends: 18s → 30s (4-concurrent and 16-concurrent cases needed 21.8–28.8s)
• realistic/payload_complexity: 10s → 15s (mixed_parts and nested_metadata exceeded budget)
• realistic/connection: 10s → 15s (new_client_per_request TCP setup cost)
• enterprise/client_interceptors: 8s → 10s (5 and 10 interceptor chains exceeded budget)

Benchmark Code Improvements:

• production_scenarios.rs: Pre-allocate direct_params outside the measurement loop in bench_dispatch_routing to isolate handler dispatch cost from fixture allocation, producing fairer comparison against HTTP round-trip path
• Added detailed measurement notes to data_volume.rs explaining BTreeSet rebalancing variance in the after_prefill/10000 case

Documentation Updates:

• Updated README.md to reference all 237 benchmarks across 13 suites
• Updated .github/workflows/README.md to document the benchmarks workflow

Notable Details

• All timeout increases are based on actual CI analysis showing which benchmarks were marginally exceeding their budgets
• The dispatch routing benchmark fix isolates the true handler cost by moving fixture allocation outside the timed loop
• Documentation entries explain why seemingly anomalous results (e.g., direct handler slower than HTTP) are actually expected and validate the measurement methodology

https://claude.ai/code/session_01YNLEZUUKzzs4qjkj1TPgXz