feat(cpp): performance benchmarks and binary size tracking

[itomek]

Closes #358

Summary

• Adds benchmark_cpp.yml workflow measuring binary size, startup time, loop latency, and memory footprint across Linux and Windows
• Baselines are cached on pushes to main; every PR run compares current results against the last baseline with per-metric thresholds (10% for binary size, 15% for everything else)
• Regression report is printed inline to the CI log; results uploaded as downloadable artifacts

Fixes applied (from CI code review)

• Cache path mismatch — restore path now matches save path; renamed to benchmark-baseline.json before the fresh run overwrites benchmark-results.json
• CI trigger — benchmark_cpp.yml and build_cpp.yml added to both push and pull_request paths in build_cpp.yml
• Windows header order — <windows.h> before <psapi.h> in bench_utils.h
• Benchmark job timeout — timeout-minutes: 15 added
• Double-init guard — initCalled_ bool prevents duplicate tool registration in BenchAgent
• Missing metric detection — compareAndReport now flags baseline metrics absent from current run as MISSING (catches crashed benchmarks that previously produced a false-positive clean pass)
• Timer guard — elapsedUs() returns 0 if stop() was never called
• CMakeLists option order — GAIA_BUILD_BENCHMARKS moved to top-level options block alongside other GAIA_BUILD_* options
• .gitignore — cpp/benchmark-*.json excluded (ephemeral CI artifacts)

Test plan

• Built locally (clean, no new warnings)
• Ran benchmarks end-to-end — JSON output verified
• Ran compare mode — regression table printed, exit 1 on detected regression
• 216/217 unit tests pass (1 pre-existing failure unrelated to this PR — LemonadeClientTest.DefaultConstruction fails due to GAIA_CPP_BASE_URL env var set in local environment)
• Lint passed (uv run python util/lint.py --all)