mace-freeze merge

[7radians]

The addition of frozen transfer learning functionality (freezing layers/parameter tensors) for fine-tuning of foundation models. Please refer to the updates in README.md file for details, and cite https://arxiv.org/abs/2502.15582

[7radians]

Fixed AttributeError: 'NoneType' object has no attribute 'layout' in the check gradients function, the pre-commit tests look fine on my machine, hopefully no more issues

[7radians]

The latest addition is learning rate rescaling, which acts as a 'soft-freeze' mode. It assigns a low learning rate to certain parameters instead of freezing them completely. This can be used alongside freezing as a buffer (frozen par → soft-frozen par → active par), with multihead fine-tuning (for the deeper layers), or on its own. I'll add instructions to the README soon, along with some tests.

[7radians]

*this update is only compatible with cueq 0.4.0, the earlier cueq versions with mace 0.3.13 do not handle None grads