feat: add bond length distribution benchmark

[lwalew]

Pre-review checklist for PR author

PR author must check the checkboxes below when creating the PR.

• I've confirmed the contribution guidelines.

Summary

Migrates the bond length distribution benchmark from the MLIP Audit suite. For each of a set of small organic molecules — covering the C-C, C=C, C#C, C-N, C-O, C=O and C-F bond types — an NVT molecular dynamics simulation is run at 300 K starting from a QM-optimised reference geometry (selected from QM9), and the deviation of a tracked covalent bond from its reference length is measured over the trajectory. A well behaved potential keeps bonds close to their reference length, so a lower average deviation is better.

The analysis reports the average bond length deviation per model in the metrics table and shows the distribution of sampled deviations as a histogram.

Linked issue

Resolves #671

Progress

• Calculations
• Analysis
• Application
• Documentation

Testing

Not yet run end-to-end: the benchmark input data still needs to be uploaded to the ML-PEG S3 bucket at inputs/molecular_dynamics/bond_length_distribution/bond_length_distribution.zip (layout bond_length_distribution/bond_length_distribution.json). Kept as a draft until the data is available.

Verified locally in the meantime:

• pre-commit passes on all changed files (ruff check, ruff format, numpydoc-validation).
• Both test modules are collected by pytest with mlipaudit installed, and skip cleanly with Please install mlipaudit extra when it is not, rather than erroring during collection.
• The element info written to info.json was checked against the dataset: 8 molecules, and the union of the per-molecule element lists is ['C', 'F', 'H', 'N', 'O'], matching BondLengthDistributionBenchmark.required_elements.
• The calculation → analysis hand-off was exercised end-to-end with a local copy of the dataset, confirming the analysis reads the dataset from the calculation outputs without downloading it again.

New decorators/callbacks

No new callbacks required; uses the existing plot_hist and build_table decorators and the plot_from_table_column callback.

Note for reviewers

Addressing the review feedback here and the equivalent points on #658:

• Element filtering. Follows the Folmsbee template, with elements saved as one list per molecule rather than a single union, so partial filtering can exclude individual molecules later. Ordering follows the dataset, which matches the molecule order in analyze()'s results. This matters here: only 1 of the 8 molecules contains F, so under a single union list, filtering out F would exclude the whole benchmark rather than one molecule.
• Reusing the calculation output. The calculation now copies the downloaded dataset into its outputs, and the analysis reads it from there with data_input_dir=CALC_PATH, so it no longer re-downloads from S3 (previously twice per run). A missing dataset raises a clear "Please run the calculation" error.
• pytest.importorskip. Added to both the calculation and analysis modules, positioned ahead of the ml_peg.calcs.utils.mlipaudit import, since that module imports mlipaudit unconditionally.
• GPU cost estimate. Added to the docs, per the comment thread below.
• Unlike Folmsbee and tautomers, this benchmark does not use the mock output as its source of truth. Those need it to order parity-plot hoverdata labels; the histogram here has no per-molecule labels, so the dataset in the calculation outputs is sufficient.
• A model whose simulations all fail is reported as NaN rather than None.

[joehart2001]

gpu estimate?

[lwalew]

There are 8 molecules of size 4-13 atoms, each doing 1M steps so for instance for UMA-S's throughput of 1.4ns/day/1000 atoms on an H100, we can expect at least 2h for instance. Although there will be some overhead associated with the simulation engine so for most models, a couple of hours is realistic.

[joehart2001]

Thanks for the pr! same as #658 could you send the data + have a look at the element filtering?

docs: https://ddmms.github.io/ml-peg/developer_guide/filter.html