feat: add protein sampling benchmark

[lwalew]

Pre-review checklist for PR author

• I've confirmed the contribution guidelines.
• I have reviewed and understand all AI-generated code in this PR.
• I have added human-written tests for the new logic.
• I have properly cited any upstream algorithms or libraries the AI utilized.
• I have disclosed significant AI tool usage in the PR description.

Summary

Migrates the protein sampling benchmark from the MLIP Audit suite into ml-peg,
introducing a new top-level biomolecules category.

The benchmark runs short molecular dynamics simulations for a set of small proteins
(chignolin 1UAO, Trp-cage 2JOF, orexin beta 1CQ0), collects the sampled backbone
(phi/psi) dihedral angles, and compares their distributions to reference
distributions. Three metrics are reported (all lower-is-better, averaged over residue
types and stable systems):

• Backbone Dihedral RMSD — RMSD between sampled and reference backbone dihedral
  distributions.
• Backbone Hellinger Distance — Hellinger distance between the same distributions.
• Backbone Outliers Ratio — fraction of sampled backbone dihedrals lying far from
  any reference data point.

This is a table-only benchmark (no primary plot), consistent with the migration spec.

The underlying SamplingBenchmark logic lives in the pinned
MLIP Audit package (mlpeg-migration
branch); ml-peg wires it up through a thin MlPegSamplingBenchmark adapter that
disables skip_if_elements_missing (ml-peg's ASE calculators do not expose the set of
supported elements).

Notes for reviewers

• This PR adds the new biomolecules category (app YAML, docs category page, docs
  toctree entry).
• It also adds the shared mlipaudit wiring (pyproject.toml optional extra +
  conflict + tool.uv.sources, and the mlip_audit framework entry in
  frameworks.yml). This wiring overlaps with other in-flight MLIP Audit migration
  PRs (e.g. ring planarity, protein folding) and is expected to be reconciled at
  merge.

Linked issue

Resolves #726

Progress

• Calculations
• Analysis
• Application
• Documentation

Testing

Static/local validation only so far: all new Python modules compile and the full
pre-commit suite passes (ruff-check, ruff-format, numpydoc-validation, etc.).
End-to-end runs across models are pending upload of the input data to S3
(inputs/biomolecules/protein_sampling/protein_sampling.zip), which has not been
performed as part of this PR.

New decorators/callbacks

None. Uses the existing build_table decorator and BaseApp table-only layout.