Remove python level MMFF prop shim in favor of direct C++

[scal444]

The weird python cache wasn't actually working for all use cases, which was being hidden in unit tests, unfortunately, due to the call to capture_mmff_settings.

The difficulty in general was that this class isn't exposed via public API to the python interface on the rdkit side, but since we can see it, we can create our own quick wrapper.

[greptile-apps]

Greptile Summary

This PR removes the Python-level WeakKeyDictionary MMFF settings cache (which silently failed for externally-configured MMFFMolProperties objects) and replaces it with a C++ helper, buildMMFFPropertiesFromRDKit, that reads settings directly from the underlying RDKit::MMFF::MMFFMolProperties via a layout-compatible shim. The approach is well-documented, the pre-existing ODR/ABI fragility is acknowledged with warnings and a TODO pointing at the upstream RDKit issue, and a targeted new test demonstrates the fix works for the exact failure mode described.

Confidence Score: 5/5

Safe to merge — fixes a real bug, no new correctness issues introduced, pre-existing platform limitations are clearly documented

No P0 or P1 issues found. The ODR and ABI fragility concerns raised in the previous review thread have been addressed with explicit \warning docs, a static_assert, and a TODO tracking the upstream fix. The new test directly exercises the formerly broken code path.

No files require special attention

Important Files Changed

Filename	Overview
nvmolkit/mmff_python_utils.h	Adds ForceFields::PyMMFFMolProperties shim and buildMMFFPropertiesFromRDKit; relies on Linux ODR symbol merging (documented, previously reviewed, with static_assert guard and TODO)
nvmolkit/_mmff_bridge.py	Removes the fragile Python-level WeakKeyDictionary settings cache and delegates directly to _batchedForcefield.buildMMFFPropertiesFromRDKit; clean simplification
nvmolkit/batchedForcefield.cpp	Adds Boost.Python binding for buildMMFFPropertiesFromRDKit with correct arg names and docstring
nvmolkit/tests/test_batched_forcefield.py	Removes capture_mmff_settings dependency; adds test_mmff_batched_forcefield_reads_externally_configured_properties covering the exact bug the PR fixes
nvmolkit/tests/test_mmff_optimization.py	Removes capture_mmff_settings import; replaces capture_mmff_settings call with a direct return mmff_props

_{Reviews (3): Last reviewed commit: "Add explicit support test" | Re-trigger Greptile}

[evasnow1992]

Changes look good to me. Thank you for fixing this bug. May benefit from adding one test for testing building properties with raw rdForceFieldHelpers.MMFFGetMoleculeProperties(mol), then configuring with Set*Term() calls, and handing to MMFFBatchedForcefield, to see if the previous broken path works now. Not a blocker.