Expose FIRE minimizer Python APIs

[scal444]

PR 6 for #216

[greptile-apps]

Greptile Summary

This PR exposes the FIRE energy minimizer as a first-class Python API for both the standalone (MMFFOptimizeMoleculesConfs, UFFOptimizeMoleculesConfs) and batched-forcefield (MMFFBatchedForcefield.minimize, UFFBatchedForcefield.minimize) code paths. It introduces a new _types Boost.Python module that wraps FireOptions, re-exports it through types.py, and wires it into all four minimization entry points with Python-level validation and C++ routing via a local MinimizerKind enum.

• A new _types module (types.cpp / _types.pyi) exposes FireOptions with all tuneable fields; types.py re-exports it as a public symbol.
• minimizerKind and fireOptions parameters are added to the MMFF and UFF standalone functions and the shared _BatchedForcefieldBase._minimize() method, with Python-level validation before forwarding to C++.
• Both stubs for the *ConfsDevice variants (_mmffOptimization.pyi and _uffOptimization.pyi) mark several required C++ parameters as optional — those args have no Boost.Python default, so calling the private C++ module directly without them will raise a runtime error.

Confidence Score: 5/5

The public Python API and C++ routing logic are correct; the only rough edges are in the private _mmffOptimization and _uffOptimization stub files.

All changes to the public-facing Python wrappers and C++ bindings are consistent and well-tested. The stub inaccuracies affect only private underscore modules that end-users are not expected to call directly, and the Python wrappers always supply all required arguments explicitly before delegating to those modules.

nvmolkit/_mmffOptimization.pyi and nvmolkit/_uffOptimization.pyi — the newly added ConfsDevice stubs show required C++ parameters as optional.

Important Files Changed

Filename	Overview
nvmolkit/types.cpp	New Boost.Python module exposing FireOptions struct; all fields correctly wired with def_readwrite.
nvmolkit/_types.pyi	New stub for _types module; all FireOptions fields listed, __init__ correctly annotated.
nvmolkit/types.py	Re-exports FireOptions from the new _types C++ module via noqa: F401; minimal and correct.
nvmolkit/_mmffOptimization.pyi	New MMFFOptimizeMoleculesConfsDevice stub incorrectly marks maxIters, properties, hardwareOptions, and targetGpu as optional; those are required in the C++ binding.
nvmolkit/_uffOptimization.pyi	New UFFOptimizeMoleculesConfsDevice stub marks five required C++ args as optional; fireOptions: object is inconsistent with the MMFF stub (Any).
nvmolkit/mmffOptimization.cpp	Adds FIRE path to both MMFFOptimizeMoleculesConfs and MMFFOptimizeMoleculesConfsDevice; backend/minimizer parsing extracted into local helpers; C++ defaults added to MMFFOptimizeMoleculesConfs binding.
nvmolkit/uffOptimization.cpp	Mirrors MMFF changes for UFF; FIRE path added; hardwareOptions given a default in the UFFOptimizeMoleculesConfs binding.
nvmolkit/batchedForcefield.cpp	FIRE path added to all four minimize methods; fireOptionsWithGradTol helper always overwrites gradTol with the forceTol parameter (previously flagged).
nvmolkit/batchedForcefield.py	Adds minimizerKind / fireOptions to _BatchedForcefieldBase._minimize() and propagates them through both MMFFBatchedForcefield and UFFBatchedForcefield overloads with proper validation.
nvmolkit/mmffOptimization.py	Adds backend, minimizerKind, fireOptions to the public MMFFOptimizeMoleculesConfs wrapper with Python-level validation and alias normalisation before forwarding to C++.
nvmolkit/uffOptimization.py	Adds minimizerKind / fireOptions to UFFOptimizeMoleculesConfs; no backend parameter (UFF FIRE uses a fixed backend internally, asymmetric with MMFF but consistent with the C++ binding).
nvmolkit/tests/test_mmff_optimization.py	Adds FIRE correctness tests (BATCHED and PER_MOL backends) plus validation tests for unknown minimizerKind and backend; coverage looks adequate.
nvmolkit/tests/test_uff_optimization.py	Adds FIRE correctness test and minimizerKind rejection test for the UFF path; matches the MMFF test structure.
nvmolkit/tests/test_batched_forcefield_device.py	New test_minimize_fire_host_returns_nested_energy_and_status covers MMFF and UFF batched-forcefield FIRE path; checks shape and type of returned tensors.
nvmolkit/tests/test_types.py	New test_fire_options_exposes_native_defaults_and_mutators verifies default gradTol and field mutability; straightforward and correct.
nvmolkit/CMakeLists.txt	Adds _types CMake target for the new types.cpp module; structure matches existing targets.

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

[greptile-apps]

forceTol silently overrides user-supplied fireOptions.gradTol

In the C++ binding (batchedForcefield.cpp), the FIRE path calls fireOptionsWithGradTol(fireOptions, gradTol), which copies the user's FireOptions and overwrites its gradTol with the forceTol parameter (default 1e-4). Any explicit fireOptions.gradTol the caller sets is therefore silently discarded. The new test illustrates this: it sets options.gradTol = 1e-3 and calls minimize(maxIters=100, minimizerKind="FIRE", fireOptions=options) — but forceTol defaults to 1e-4, so the test is unknowingly running with a different tolerance than intended. Neither the minimize() docstring nor the FireOptions.gradTol field documents this override.