pipelines: align RNG interface with RNGRoot convention

[shaypal5]

Context

Pipeline functions in leadforge/pipelines/build_v5.py accept np.random.RandomState directly. The rest of the package standardizes on leadforge.core.rng.RNGRoot with named substreams for deterministic derivation.

Now that these functions are in-package, they should follow the same convention.

What to do

1. Change pipeline function signatures to accept a seed (int) and derive per-step sub-RNGs via RNGRoot, OR accept an RNGRoot and call .child("subsample") etc.
2. Update scripts/build_v5_snapshot.py orchestration to pass seed/RNGRoot instead of raw RandomState
3. Verify determinism is preserved (same seed → same output)
4. Update tests accordingly

Considerations

• np.random.RandomState is legacy; np.random.Generator is preferred for new code
• RNGRoot currently wraps random.Random — may need a numpy substream method

Origin

Copilot review comment on PR #30.