Add a BatchRelaxer to speed up geometry optimisation of many small structures

[zhubonan]

Add a batch relaxer which combines multiple atoms as a single batch to the underlying model. This can give speed-ups when lots of smaller structures needs to be relaxed using GPU (typical while using the foundation models for high-throughput studies).

The current code is working, but a few useful functionalities can still be added:

• Support for ensemble models
• Support other types of MACE like EnergyDipoleMACE?
• Since we already run the optimiser step-by-step, it does no harm to add some detection for pathological structures (exploded/collapsed) which can make the main worker process to stuck.
• Automatically determine the best batch size by collecting the timings of the compute method.

[ilyes319]

Hey looks great, ping me when you think it is ready to merge.