Add reproducible latency benchmarks

[haasonsaas]

Context

Formal.ai claims sub-10ms p50 proxy overhead. Gate has no published latency numbers. Without benchmarks, Gate cannot compete in enterprise evaluations where latency is a hard requirement.

Proposal

Create a reproducible benchmark suite:

• Benchmark script in scripts/benchmark.sh or internal/benchmark/
• Scenarios: direct connection vs Gate-proxied for each protocol
  • PostgreSQL: simple SELECT, complex JOIN, INSERT batch
  • MySQL: equivalent queries
  • With and without policy evaluation (0 policies, 1 policy, 10 policies)
  • With and without masking enabled
• Metrics: p50, p95, p99 latency; throughput (queries/sec); CPU/memory overhead
• Hardware baseline: document the test hardware (e.g., M4 MacBook Pro, or 4-core EC2 instance)
• Automated: runs in CI on demand (not every push) to catch regressions
• Published: results in README or dedicated benchmarks page

Acceptance Criteria

• Benchmark script is reproducible and documented
• Covers PostgreSQL and MySQL with varying policy loads
• Reports p50/p95/p99 latency and throughput
• Results published with hardware specification
• Can be run in CI for regression detection