docs(benchmarks): add v0 starting-point design spec

[lsetiawan]

Summary

• Adds a new design document scoping a minimal, trustworthy v0 for the HPyX benchmarking and profiling subsystem, before the Python API surface expands further.
• Complements (does not replace) the existing benchmarking_profiling_sdd.md, which remains the long-term roadmap. This spec explicitly defers pyperf, asv, CI jobs, macro/memory suites, and numeric targets to that companion doc.

Changes

• New: docs/design-docs/2026-04-24-benchmarking-v0-design.md
  • Section 5: 7-rule benchmark authoring contract (no timed setup, 3-order-of-magnitude size sweep, required NumPy + pure-Python + concurrent.futures baselines in the same group, explicit group names, minimize incidental Python overhead, thread-scaling parametrization, free-threading gating).
  • Section 6: shared benchmarks/conftest.py fixtures (pin_cpu, seed_rng, no_gc, hpx_runtime, hpx_threads, requires_free_threading, env_sanity_check) with an explicit resolution for the session-runtime vs. thread-count conflict.
  • Section 7: profile CMake preset (RelWithDebInfo + -fno-omit-frame-pointer, LTO off, default visibility) so py-spy --native, perf, and memray --native can resolve C++ frames.
  • Section 8: scripts/run_bench_local.sh subcommands and a one-page contributor benchmarks/README.md.
  • Section 9: planned updates to the companion SDD (directory rename bench/ → benchmarks/, elevate Python 3.13 free-threading to a first-class axis, prepend "Phase 0" to the rollout plan).
  • Section 10: acceptance criteria for v0.

Test plan

• Reviewer confirms v0 scope is small enough to land in a single implementation plan.
• Reviewer agrees with the session-scoped hpx_runtime fixture + dedicated file for thread-count-varying benchmarks (Section 6.1).
• Reviewer agrees that concurrent.futures.ThreadPoolExecutor is the right third baseline alongside NumPy + pure Python.
• Reviewer agrees with deferring pyperf/asv/CI/macro/memory/TARGETS.md to the existing SDD (Section 12).
• No follow-up edits block moving to an implementation plan via the writing-plans workflow.