fix PBC on GPU bug in state_to_atoms

[cw-tan]

Summary

Bug likely appeared because PBC is not a torch.Tensor instead of just a bool (#320). This failure mode only shows up when running on GPU, which explains why CI didn't catch it (presumably it's only run on CPU). I didn't check the other format conversion utility functions, but potentially some of the others might have a similar failure mode if they were not changed with the breaking PBC internal behavior change.

  File "/home/cw/harvard/matbench-discovery/models/nequip/torchsim_evals/test_nequip_discovery.py", line 172, in <module>
    relaxed_atoms_list = ts.io.state_to_atoms(final_state)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/cw/harvard/torch-sim/torch_sim/io.py", line 67, in state_to_atoms
    atoms = Atoms(
            ^^^^^^
  File "/home/cw/micromamba/envs/nequip/lib/python3.12/site-packages/ase/atoms.py", line 254, in __init__
    self.set_pbc(pbc)
  File "/home/cw/micromamba/envs/nequip/lib/python3.12/site-packages/ase/atoms.py", line 571, in set_pbc
    self.pbc = pbc
    ^^^^^^^^
  File "/home/cw/micromamba/envs/nequip/lib/python3.12/site-packages/ase/atoms.py", line 567, in pbc
    self._pbc[:] = pbc
    ~~~~~~~~~^^^
  File "/home/cw/micromamba/envs/nequip/lib/python3.12/site-packages/torch/_tensor.py", line 1213, in __array__
    return self.numpy().astype(dtype, copy=False)
           ^^^^^^^^^^^^
TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

Checklist

Before a pull request can be merged, the following items must be checked:

• Doc strings have been added in the Google docstring format.
• Run ruff on your code.
• Tests have been added for any new functionality or bug fixes.

[curtischong]

This looks fine. Thanks for catching this. We should probably run some of our tests on gpu