Add UFF forcefield C++ implementation

[scal444]

Adds UFF bond stretch, angle bend, torsion, inversion, and van der Waals terms with batched GPU evaluation. Includes RDKit UFF parameter flattening, batched forcefield wrapper, and C++ tests.

[greptile-apps]

Greptile Summary

This PR adds a complete GPU-accelerated UFF forcefield implementation covering bond stretch, angle bend, torsion, inversion, and van der Waals terms, along with a batched forcefield wrapper and RDKit parameter flattener. The energy and most gradient kernels are mathematically sound and well-tested against RDKit reference values.

Key findings:

• Inversion gradient sign error (uff_kernels_device.cuh): uffInversionGrad constructs the JI–JK plane normal via crossProduct(-rJIx, ..., rJKx, ...) (negated rJI), while uffCalculateCosY (used in the energy) uses positive rJI. After normalisation the gradient's cosY equals -sinW rather than +sinW. The dE_dW formula does not account for this sign flip, producing K·sinW·(C1 − 4C2·cosW) instead of the correct K·(C1 − 4C2·sinW). The error is proportional to the out-of-plane displacement, so it is small for typical near-planar sp2 geometries (which is why the current gradient tests likely pass within the 2×10⁻⁴ tolerance), but can be a factor-of-2 or larger error for significantly non-planar centres.
• Three pre-existing issues acknowledged in comments: zero-distance bond gradient fallback, VdW gradient/energy inconsistency at dist ≤ 0, and the UFF→MMFF header coupling — all tracked with TODOs.
• Energy kernels, torsion and angle gradient derivations, VdW LJ gradient formula, and Chebyshev polynomial implementations are all correct.
• The block-per-molecule shared-memory gradient kernel correctly caps at 256 atoms and falls back to global memory for larger systems.

Confidence Score: 4/5

• Safe to merge with caution — one P1 inversion gradient bug will produce incorrect gradients for non-planar sp2 centres, affecting minimisation quality on such molecules.
• The bulk of the implementation (energy terms, torsion/angle/VdW gradients, batching infrastructure, build system) is correct and well-tested. However, the inversion gradient sign error is a definite present defect: the plane normal direction used in the gradient computation is inverted relative to the energy function, yielding a systematically wrong derivative for non-planar inversion centres. The current tests are unlikely to catch this because they compare at near-planar equilibrium geometries where the inversion contribution is inherently small. This warrants a 4/5 rather than 5/5.
• src/forcefields/uff_kernels_device.cuh — specifically uffInversionGrad around the crossProduct call and the dE_dW formula.

Important Files Changed

Filename	Overview
src/forcefields/uff_kernels_device.cuh	Core device-side kernel implementations for all UFF energy/gradient terms; contains a P1 inversion gradient bug where the plane normal used in uffInversionGrad is negated relative to uffCalculateCosY, causing incorrect gradient magnitudes for non-planar sp2 centres.
rdkit_extensions/uff_flattened_builder.cpp	Flattens RDKit UFF parameters into SoA host buffers; logic closely mirrors RDKit's UFF builder. The unusual neighborSlot increment pattern in addInversions (skipping slot 1 for the central atom) is correct but hard to follow.
src/forcefields/uff.cu	Host-side batch management: stream assignment, host-to-device transfers, addMoleculeToBatch, and kernel dispatchers for energy and gradient. Implementation is consistent and handles all term types including constraint terms borrowed from MMFF.
src/forcefields/uff_kernels.cu	Kernel launchers and two block-per-molecule kernels (combined energy + gradient). The blockSizePerMol = 128 shared-memory gradient path correctly limits to 256 atoms; fallback to global memory is correct.
tests/test_uff.cu	Validates GPU UFF against RDKit reference for energy, gradient, batched, and minimizer paths. Tests cover edge cases (single-atom SMILES, small molecules) and a 25-molecule validation set. Gradient tolerance of 2e-4 may be too loose to catch the inversion gradient bug on near-planar test molecules.

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

[scal444]

Partial implementation for #114

has not yet been optimized, we'll go through the kernels but lower priority.

[evasnow1992]

Changes look good to me. Understand that further optimizations may be applied later. Also a note that python APIs haven't been added yet and document files like docs/index.rst may need updates to include UFF alongside MMFF.