feat: add inference speed 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 inference speed benchmark from the MLIP Audit suite into ml-peg under the molecular_dynamics category.

For each structure in a size-stratified protein dataset (elements N, H, O, S, C), the underlying InferenceSpeedBenchmark times the model forward pass (energy + forces) and runs a short MD simulation per backend. This PR wires that up for ml-peg's ASE calculators and exposes:

• Table metric Forward Time / Atom (µs/atom, lower is better): mean over structures of average_forward_time / num_atoms, skipping structures whose forward pass failed.
• Scaling plot (line plot): forward-pass time (s) vs. number of atoms, one line per model, sorted by system size and restricted to successful forward passes.

This is a wall-clock, hardware-dependent measurement (H100 reference), not a level of theory, so results are only comparable across models run on the same hardware.

⚠️ Blocked

This PR is blocked and must remain a draft until the inference_speed benchmark lands on the mlipaudit mlpeg-migration branch.

• ml-peg pins mlipaudit to the mlpeg-migration branch (it carries the ASE BaseCalculator fix and other changes the rest of ml-peg needs). That pin is intentionally left unchanged here.
• InferenceSpeedBenchmark currently exists only on the mlipaudit branch feat/scaling-benchmark-revamp, at src/mlipaudit/benchmarks/inference_speed/inference_speed.py. It is not on mlpeg-migration.
• Until inference_speed is merged into mlpeg-migration, the import from mlipaudit.benchmarks.inference_speed.inference_speed import ... will not resolve, so the benchmark cannot run.

Additional notes:

• The shared mlipaudit wiring (the mlipaudit optional extra + conflicts entry in pyproject.toml, the [tool.uv.sources] pin, and the mlip_audit entry in frameworks.yml) is already present on main and overlaps with other in-flight mlipaudit PRs, so this PR does not touch those files.
• Data packaging requirement: the S3 object inputs/molecular_dynamics/inference_speed/inference_speed.zip must contain an inference_speed/ directory of size-prefixed .xyz files (e.g. 121_1ay3.xyz) — the benchmark reads data_dir / "{structure}.xyz" from data_input_dir / "inference_speed".

Linked issue

Resolves #742

Progress

• Calculations
• Analysis
• Application
• Documentation

Testing

Static checks only so far: python -m py_compile on the new modules and all pre-commit hooks (ruff-check, ruff-format, numpydoc-validation, whitespace/EOF) pass.

End-to-end testing is pending two prerequisites:

1. The mlipaudit dependency landing on mlpeg-migration (see Blocked above).
2. The S3 data upload of inputs/molecular_dynamics/inference_speed/inference_speed.zip.

The good/bad thresholds in metrics.yml (0.5 / 5.0 µs/atom) and the score midpoint are estimates and need tuning against real H100 runs.

New decorators/callbacks

None. Reuses the existing @plot_scatter (line plot), @build_table, and plot_from_table_column / read_plot patterns.