ASE Calculator interface

[laserkelvin]

This PR adds a Calculator implementation that specializes to matsciml's pipeline/abstractions.

Currently it is only implemented for energy and force prediction, but given that ase also supports other properties like stress, multipole moments, etc. there is some flexibility in using ScalarRegressionTask for things like that.

• Add matsciml.interfaces.ase.MatSciMLCalculator class that adapts Atoms representations to structures expected by matsciml models
• Add two minimum working examples for using the new interface: one loading from an inexistent pretrained checkpoint, and another showing training on LiPS and using ase after a quick training run on the devset
• Adds a Github action to run pytests in the interfaces submodule
• Removes obsolete OCP relaxation and ase routines

[melo-gonzo]

Looks good so far. Left a few comments.

One more general comment is that it took me a while to figure out that using a checkpoint which was created from a task like this:

model = EGNN(encoder_kwargs=...)
task = ForceRegressionTask(
    model,
    output_kwargs=...
)

will not load properly, however a checkpoint created from a task like this:

task = ForceRegressionTask(
    encoder_class=EGNN,
    encoder_kwargs=...,
    output_kwargs=...
)

will load. This is due to how the 'hyper parameters' are stored in the model checkpoint.

It could be worth depreciating the first method of task creation in matsciml, or making note of this somewhere.

[melo-gonzo]

Can you consolidate this between the class methods?

[laserkelvin]

You mean the path conversion?

[melo-gonzo]

This bit should be consolidated to one function:

        if isinstance(ckpt_path, str):
            ckpt_path = Path(ckpt_path)
        if not ckpt_path.exists():
            raise FileNotFoundError(f"Checkpoint file not found; passed {ckpt_path}")

[laserkelvin]

Addressed in 439b7b4