feat: add LoadAtomsDataModule for load-atoms datasets

[manishpaulish]

Adds LoadAtomsDataModule, a Lightning DataModule that wraps the
load-atoms package to provide
easy access to its dataset collection directly from the NequIP training infrastructure.

Any dataset available via load_atoms.load_dataset() (e.g. C-GAP-17, QM9,
rMD17-aspirin) can be used without manual downloading or format conversion.

The implementation:

• Downloads and caches the dataset via load-atoms in prepare_data()
• Writes to extxyz format for consumption by the existing ASEDataset pipeline
• Supports the standard train_val_test_split, subset_len, transforms,
  include_keys, exclude_keys, and key_mapping arguments consistent with
  other DataModules in this codebase

This was suggested by @cw-tan in #600.

[cw-tan]

Hi thanks for the contribution. The first thing I'd flag is that we shouldn't make load-atoms a mandatory dependency. Users who want to use the datamodule should install it themselves, and instruction to install them should be in the docstring. The new datamodule should also be exposed in the docs.

[manishpaulish]

Thanks for the feedback. Removed load-atoms from mandatory dependencies and added an install note to the docstring. Also exposed the class in docs/api/datamodule.rst.

[cw-tan]

Thanks @manishpaulish . Have you verified that the datamodule works fine when used in a typical traning config for several examples of load-atoms datasets? Also, have you built the docs to check that there are no errors?

[manishpaulish]

Verified the datamodule works correctly with C-GAP-17 and C-GAP-20U (both download, load, and batch correctly). Docs build succeeds with make html. The only warning is a pre-existing one about torchsim.NequIPTorchSimCalc, unrelated to this PR.