feat(mlip): ALCHEMI Polaris benchmark harness + batched MD driver

[tdpham2]

Summary

Adds two capabilities on the existing nvalchemi-mace (NVIDIA ALCHEMI) MLIP backend:

1. A Polaris throughput benchmark harness — the repo previously had only smoke.py (correctness/integration evidence, explicitly not a benchmark).
2. A batched MD driver (driver="md") — MD previously existed only in uma.py (FAIRChem, single-structure). ALCHEMI's batched path handled only energy/opt.

Benchmark harness — alcf/polaris/mlip/benchmark/

• bench.py — single-GPU: batch-size sweep (structures/sec + peak GPU mem via nvidia-smi), structure-size scaling (supercells), and ALCHEMI-vs-ase-mace baseline. Emits bench_results.jsonl + bench_meta.json.
• bench_multigpu.py — node-level throughput: one process per GPU (CUDA_VISIBLE_DEVICES pinned) over round-robin-sharded inputs, aggregated into node_result.json. No runner change needed (runner is single-GPU per process).
• bench.pbs / bench_multigpu.pbs — Polaris PBS scripts modeled on smoke.pbs (conda module, ALCF proxy, model-weight pre-warm before timed runs).
• README.md — how to run each sweep, read results, and hardware/parity caveats.

Batched MD driver

• config.py: MLIPCalculationConfig gains ensemble, temperature, timestep, md_steps, friction with validation; driver extended to include md; option gating updated in _validate_explicit_options.
• cli.py: --driver md and --ensemble/--temperature/--timestep/--md-steps/--friction on the shared _mlip_options; MD restricted to nvalchemi-mace.
• runner.py: MD branch in _run_nvalchemi_chunk, with the actual integrator construction isolated behind _NVALCHEMI_MD_INTEGRATORS + _nvalchemi_md_dynamics, plus n_steps reporting and MD-only-on-ALCHEMI validation.

⚠️ One Polaris-only step remains

The real nvalchemi.dynamics MD integrator class names/kwargs cannot be verified off-GPU (the package installs only with a CUDA extra). The code assumes VelocityVerlet (NVE) / Langevin (NVT) with (model, hooks, dt, n_steps, temperature[, friction]). On Polaris, run python -c "import nvalchemi.dynamics as d; print(dir(d))" and, if they differ, update the single isolated mapping in runner.py — a clear error is raised if the assumed names are absent.

Testing

• Added tests: MD dispatch/result-mapping (test_mlip_nvalchemi.py), config validation (test_mlip_validation.py), CLI gating (test_cli.py).
• Full MLIP/CLI suite green locally (only the unrelated pre-existing test_mcp_api.py import error was excluded).
• Benchmark scripts validated for imports, arg-parsing, and structure prep; GPU runs are Polaris-only.

🤖 Generated with Claude Code