perf, establish MVP bench suite for I/O throughput claims

[steven-varga]

Goal

Replace hand-wavy performance claims with reproducible, versioned numbers. See design report: tasks/h5cpp-performance-evaluation-framework-design.md in workspace.

MVP Scope

1. Vendored nanobench harness (single header, validated in libdecimal)
2. bench/ directory with CMake integration (H5CPP_BUILD_BENCH)
3. 5 bench groups:
   • sequential write/read (1D/2D/3D vectors, matrices)
   • container overhead (raw HDF5 C vs h5cpp high-level)
   • chunked + filtered write (gzip, zstd, lz4, shuffle)
   • h5::high_throughput speedup (direct chunk vs standard)
   • small-object latency (single struct, 4KB vector)
4. Filesystem baselines: raw POSIX fwrite, mmap, memcpy, tmpfs
5. JSON + console output
6. CI job on ubuntu-24.04/gcc-14 (advisory, not merge gate)
7. bench/README.md

Out of Scope

• MPI (deferred to Phase 3)
• S3 backend (greenfield, does not exist yet)
• Threaded pipeline (blocked on implementation)
• macOS/Windows CI (Linux is stable reference)

Target Branch

staging

---

Design report authored by Winston (Architecture), John (Product), Mary (Business Analysis).