Improve indexation of the cummulated_loss_tensor in the Trainer class

[le1nux]

As seen here,

modalities/src/modalities/trainer.py

Line 144 in a599005

post_processing_fun=lambda t: torch.stack([t[0] / t[-1], t[1] / dist.get_world_size()]),

, the cumulated_loss_and_gradient_norm is indexed directly in the train function, adding unnecessary complexity to the code.

A better solution would be to encapsulate the cumulated_loss_and_gradient_norm within a dedicated class that implements also the reduce operations.