Add custom options for MMFF optimization

[scal444]

Allows better control over dielectric, interfrag interactions, and nonbond interaction limit. Partial implementation of #70

[greptile-apps]

Greptile Summary

This PR extends MMFFOptimizeMoleculesConfs with fine-grained control over the MMFF force field: callers can now supply an MMFFMolProperties object (or a per-molecule list), a scalar or per-molecule nonBondedThreshold, and ignoreInterfragInteractions. The Python–C++ bridge pattern correctly extracts only the configuration settings from the RDKit properties object (variant, dielectric, term toggles) and lets the C++ side derive all molecule-specific atom-type data from the ROMol* directly, so broadcasting a single properties object across multiple molecules is intentional and safe.

Key changes:

• mmffOptimization.py — three new public parameters with _normalize_scalar_or_list / _normalize_properties helpers that handle scalar-vs-sequence and single-object-vs-per-molecule cases.
• mmffOptimization.cpp — extractMMFFProperties / extractMMFFPropertiesList unpack the Python transport struct into nvMolKit::MMFFProperties; the old scalar nonBondedThreshold argument is fully replaced by the properties list.
• _mmffOptimization.pyi — stub updated; the declared default (None) is a minor mismatch vs the actual C++ default (empty list), flagged below.
• Tests — three well-constructed new integration tests including a fragmented-molecule fixture (now with a fixed random seed) and per-molecule property/threshold/interfrag combinations.

Confidence Score: 5/5

• Safe to merge — no functional bugs found; the single finding is a minor stub inconsistency in an internal module.
• All logic, validation, and bridging code is correct. The only issue found (P2) is that _mmffOptimization.pyi advertises None as the default for properties while the C++ binding uses an empty list. This has no runtime impact on the public API since the Python wrapper always supplies an explicit list. No P0 or P1 issues remain.
• nvmolkit/_mmffOptimization.pyi — minor stub default mismatch.

Important Files Changed

Filename	Overview
nvmolkit/mmffOptimization.cpp	Replaces scalar nonBondedThreshold with a per-molecule properties list; adds extractMMFFProperties and extractMMFFPropertiesList helpers that unpack the Python transport object into nvMolKit::MMFFProperties. Size validation is sound.
nvmolkit/mmffOptimization.py	Public API adds properties, nonBondedThreshold, and ignoreInterfragInteractions parameters; normalisation helpers correctly handle scalar-vs-sequence and single-vs-list MMFFMolProperties. Logic for broadcasting a single config object to all molecules is intentional (nvMolKit derives atom-specific parameters from ROMol* in C++, not from the properties object).
nvmolkit/_mmffOptimization.pyi	Stub updated to reflect new properties parameter, but the declared default (None) does not match the C++ binding default (empty boost::python::list()); passing None directly to the extension would raise a Boost.Python TypeError.
nvmolkit/tests/test_mmff_optimization.py	Three new tests cover custom scalar properties, per-molecule properties, and per-molecule threshold/interfrag flags. make_fragmented_mol now uses a fixed seed (42) addressing the prior review comment. make_rdkit_mmff_properties correctly calls capture_mmff_settings so the bridge can extract config settings without relying on RDKit getter availability.
nvmolkit/types.py	Trivial formatting change — two blank lines added before the _embedMolecules import. No functional changes.

_{Reviews (2): Last reviewed commit: "Format" | Re-trigger Greptile}

[evasnow1992]

Is there any specific reason we want to extract this single property instead of just passing in it as a list of one element?

[scal444]

It matters for which src cpp path we take - https://github.com/NVIDIA-Digital-Bio/nvMolKit/blob/main/src/minimizer/bfgs_mmff.h#L38-L44, but you're right we didn't need it at this level. Fixed

[evasnow1992]

Changes look good to me. Thanks!