Add mutli-GPU training, data parrallelisation, pair repulsion, distance transform and other features

[ilyes319]

Pull Request for MACE release 3.5

Description

This pull request implements a range of new features.

• [x] Multi-GPU training with distributed data parallelisation, support multi-node training,
• [x] Online data loading for large datasets,
• [x] Pair repulsion for close approach stability usign Agnesi transforms,
• [x] Distance transform for many element systems.
• [x] Interface for finetuning foundation models.
• [x] Include Chebyshev radial basis.
• [x] Compatibility of MACE with torch.compile
• [x] CLI interface.

Type of change

• [x] New feature (non-breaking change which adds functionality)
• [x] This change requires a documentation update