[Feature Request] Support batched inference with different natoms

### Summary

Currently all interface use a np.ndarray with fixed shape (nframes, natoms, 3) for coords. With dpa3 dynamic sel, one should be able to batch structures with different natoms.

### Detailed Description

I think we should support something similar to [this](https://github.com/Radical-AI/torch-sim/blob/b626d099ad5bc8c4817f76b65da80d38661cee34/torch_sim/models/interface.py#L165-L199). This need a set of entirely new interfaces from the low level descriptor to the high level DeepEval.

### Further Information, Files, and Links

_No response_