Device input/return types for MMFF optimization and Conformer generation

[scal444]

We currently re-pack all coordinates from ETKDG/MMFF/UFF back into the mol structures. A lot of workflows want to use coordinates in downstream GPU work, so they have to re-unpack and put data back on the device. It's also a slowdown between ETKDG->minimization workloads.

We should provide output APIs that are on-device for all of these, and input APIs for minimization for better chaining.

[scal444]

The big question is what format(s) we should support.

Options:

1. list/vector of [tensor], where tensor is (B, N_atoms, 3), with one tensor per molecule type.

2. Complete CSR. The tricky thing here is variable conformers per molecule, we'd need multiple levels of indexing.

3. Single padded tensor with masking (maybe multiple masks). Dims (N_mol, B_max, N_atoms_max, 3)

We could support all of these natively, or just support the CSR with python side conversion functions