feat(training): support torch losses

[JPXKQX]

Description

This PR adds a TorchLoss class to use any torch.nn class without needing to implement them inside Anemoi. This would allow us to delete the mae.py, huber.py, and the new smooth_l1.py proposed in #367 .

We would move from

training_loss:
   _target_: anemoi.training.losses.MAELoss
  scalers: ['pressure_level', 'general_variable', 'nan_mask_weights', 'node_weights']
  ignore_nans: False

to

training_loss:
   _target_: anemoi.training.losses.TorchLoss
  loss: 
     _target_: torch.nn.L1Loss
  scalers: ['pressure_level', 'general_variable', 'nan_mask_weights', 'node_weights']
  ignore_nans: False

What problem does this change solve?

This PR reduces code duplication, as it avoids the need to reimplement functionality that is already available in Torch. It also prevents the need for future PRs to test new loss functions inside Anemoi.

As a contributor to the Anemoi framework, please ensure that your changes include unit tests, updates to any affected dependencies and documentation, and have been tested in a parallel setting (i.e., with multiple GPUs). As a reviewer, you are also responsible for verifying these aspects and requesting changes if they are not adequately addressed. For guidelines about those please refer to https://anemoi.readthedocs.io/en/latest/

By opening this pull request, I affirm that all authors agree to the Contributor License Agreement.

[ssmmnn11]

Nice!

One question: what tests did you run?

[ssmmnn11]

Minor: do we write lat*lon everywhere? better gridpoints?

[mchantry]

@JPXKQX is this labelled draft for a reason?