feat: add throughput benchmark suite

[danfry1]

Summary

Adds a throughput benchmark suite so Reflow's performance is measurable and reproducible — supporting the project's "no external services, low overhead" positioning with real numbers.

bun run bench (benchmarks/index.ts) measures the engine + storage layers (trivial handlers, no-op passthrough input schema, so the numbers reflect Reflow's own cost rather than a workload):

• Enqueue throughput — createRun rate.
• End-to-end execution — enqueue → durable completion, reported as runs/sec and steps/sec.
• Across the in-memory and Bun SQLite adapters, for 1-step and 5-step workflows, at concurrency 25.

Baseline numbers (Bun, Apple Silicon) are published in the README's new Performance section. The harness warms up the JIT/disk before timing and has a drain safety-bound so it can never spin.

Base

Branched off fix/sqlite-bun-changes (#52): the Bun SQLite adapter only functions with that change-detection fix (without it, updateClaimedRunStatus returns false, runs never complete, and the benchmark stalls — which this work incidentally re-confirmed). Once #52 merges, this retargets to main cleanly (it only adds new files: benchmarks/, a bench script, and README/CHANGELOG notes).

Verified: typecheck, test, build, lint:deps, test:bun all pass; the benchmark file is independently type-clean.