Quick fix of compute_loss to be able to use fit method in tensorflow

[sararb]

Goals ⚽

The current implementation of compute_loss does not use the call methods of the Head and Model classes. This raises an issue for NextItemPredictionTask as the build method requires body to retrieve necessary information about items embeddings and masking. This class is only accessed via the build method of Head.

Implementation Details 🚧

The current PR presents a quick fix by exposing an argument call_task in the compute_loss of PredictionTask classes. The compute_loss methods of Head / Model are first computing the predictions through their call methods (to correctly build the related tasks), then computing the tasks losses with call_task=False.

Testing Details 🔍

This PR adds a test in test_model.py that defines a model with NextItemPredictionTask head, call the fit and evaluate methods, then check that losses and metrics are correctly computed.