Allow creation of MACECalculator without needing to write a checkpoint to disk.

[CompRhys]

The current MACECalculator requires you to load from a particular structure of checkpoint file. If we use a different training framework our checkpoints are not compatible and so it would be nice to be able to just pass in models.

Still needs tests.

[ilyes319]

Hey @CompRhys, can you give more context to that PR, I don't understand what you mean by checkpoint or model and training framework.

[CompRhys]

Sure, by checkpoint I mean a serialised file written to disk, by model I mean something like ShiftScaleMACE instantiated as an object, by training framework I mean something like pytorch lightning or accelerate. The checkpoint state dict for pytorch lightning cannot be loaded directly by this calculator instead requiring me to load the checkpoint access the model attribute save a compatible checkpoint to disk and then load the calculator with the path to that second compatible checkpoint. Resaving and reloading is unnecessary imo

[CompRhys]

@ilyes319 is there any pathway to this being merged?