Refactor DataSampleSchema to use tensors

[laserkelvin]

This PR refactors the DataSampleSchema to use Annotated abstraction instead of a field and model validators, and replacing the shape checking with methods in the new matsciml.datasets.validators module.

The intention behind these changes are:

1. Ensure that tensors are used first and foremost, which otherwise introduces barriers to usage in the full pipeline because tensors need to be moved to devices, expected dtypes, etc.
2. The tensor annotations in the model schema are now more semantically meaningful; e.g. CoordTensor has more implications to it than NDArray['*, 3']. For the user/developer, these effectively "named" tensors in the annotations are likely better to work with.
3. Reusable validator workflows: nominally they can be used with batched data as well.

[smiret-intel]

Looks like a good validation setup for commonly used Tensors. Can you add some brief documentation on the types for future users?

[laserkelvin]

Looks like a good validation setup for commonly used Tensors. Can you add some brief documentation on the types for future users?

They're not in the main pipeline just yet, but I've already started documenting them in the readthedocs