Expose MMFF forcefield API to python

[scal444]

Add Python bindings for MMFF batched forcefield energy and gradient
computation via Boost.Python. Includes property bridge from RDKit
MMFFMolProperties to internal MMFFProperties transport.

- NativeMMFFBatchedForcefield C++ binding with compute_energy/compute_gradients
- MMFFBatchedForcefield Python wrapper with per-molecule properties, conf_id,
  nonBondedThreshold, and ignoreInterfragInteractions support
- _mmff_bridge module for RDKit MMFFMolProperties to native conversion
- MMFFComputeEnergies/MMFFComputeGradients convenience functions
- Tests for single-mol, batched, property variants, per-molecule properties,
  and conformer id selection vs RDKit reference

[greptile-apps]

Greptile Summary

This PR exposes nvMolKit's MMFF batched forcefield to Python via a new Boost.Python extension (_batchedForcefield), a bridge module (_mmff_bridge) that converts RDKit MMFFMolProperties into nvMolKit's internal transport, a Python wrapper (MMFFBatchedForcefield) that handles scalar/per-molecule normalisation of all parameters, and a thorough test suite validating against RDKit reference values.

Key highlights:

• The previous id()-based memory-leak and phantom-reuse issue in the bridge has been addressed with weakref.WeakKeyDictionary.
• The C++ layer uses an async device-vector copy pattern with correct cudaStreamSynchronize ordering and proper CUDA error propagation.
• The extractMMFFPropertiesList fallback-to-defaults path in C++ is unreachable from the Python wrapper (which always sends exactly numMols entries), so it is purely defensive.
• The PR description lists MMFFComputeEnergies/MMFFComputeGradients convenience functions as delivered, but these do not appear in the changed files — the description should be updated or the functions added.
• The extract_mmff_settings error path does not identify the missing key by name, making debugging in exotic RDKit builds unnecessarily difficult.
• num_molecules is exposed as a public attribute alongside __len__, creating minor API duplication.

Confidence Score: 5/5

Safe to merge; all remaining findings are minor style and documentation issues with no impact on correctness.

No P0 or P1 issues were found. The prior concerns about memory leaks and id()-reuse have been addressed with WeakKeyDictionary. The three remaining comments are all P2 (missing convenience functions from PR description, a non-specific error message, and a redundant public attribute), none of which affect runtime correctness.

No files require special attention.

Important Files Changed

Filename	Overview
nvmolkit/batchedForcefield.cpp	New C++ Boost.Python extension exposing NativeMMFFBatchedForcefield with compute_energy/compute_gradients; CUDA error checking, async device-vector copy pattern, and gradient splitting all look correct.
nvmolkit/batchedForcefield.py	Python wrapper normalising scalar/per-molecule properties, conf_ids, thresholds, and lazy-building the native forcefield; minor issues: missing MMFFComputeEnergies/Gradients convenience functions and a redundant num_molecules attribute alongside len.
nvmolkit/_mmff_bridge.py	New bridge module converting RDKit MMFFMolProperties to nvMolKit's internal transport; uses WeakKeyDictionary to avoid the prior id()-based memory leak and uses fallback getters for externally-created properties objects; error message when a getter is missing could be more specific.
nvmolkit/tests/test_batched_forcefield.py	Comprehensive test suite covering single-mol, batched, property variants, per-molecule properties, and conformer id selection against RDKit reference values; tolerances and helper utilities look appropriate.
nvmolkit/CMakeLists.txt	Adds _batchedForcefield MODULE target following the same pattern as existing targets; link libraries and include directories look consistent with the rest of the build file.

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

[evasnow1992]

Thank you for exposing batch energy and gradients computation APIs to python. Changes look good to me.