fix(bench): honor explicit iteration counts

[marcus-pousette]

Why

The shared benchmark runner's small-workload sampling floor currently overrides an explicit iteration count.

That silently multiplies callers which already perform their own outer sampling. In the SQLite note-path suite, the intended 2–3 samples become 20–30 seeded database runs per result: 200 database seedings per revision while the artifact still reports only 2–3 samples. The 100-operation SQLite benchmark is multiplied similarly.

In the successful /bench run for #167, note paths consumed about 5m12s per revision. Most of that work was unreported nested sampling.

What changed

• Treat workload and environment iteration counts as explicit requests.
• Apply the automatic 7–10-sample floor only when neither supplies a count.
• Add dependency-free regression tests for workload, environment, and fallback precedence.
• Include the benchmark package test in the native Node test command so CI runs it.

Expected impact

The note-path matrix now performs its intended 20 seeded runs per revision instead of 200. Based on run 29234317596, that should save roughly nine minutes across a full base/head comparison while making the reported sample counts truthful.

Unconfigured tiny workloads retain the existing adaptive sampling behavior.

Validation

• pnpm -C packages/treecrdt-benchmark test
• Prettier check on all changed files
• git diff --check