Pass the current epoch to the aggregate function of Weighting Interface

[GiovanniCanali]

Currently, the aggregate method of any class inheriting from WeightingInterface accepts only losses, a dictionary storing the loss per condition.

In principle, this should be enough. However, some weighting schemes require updating the weights, which can be computationally expensive -- for instance, self-adaptive weighting as described in "Simulating Three-dimensional Turbulence with Physics-informed Neural Networks" (see PirateNetwork). In such cases, it may be preferable to perform these computations only every k epochs using a simple modulus operation.

Unfortunately, this is not currently possible: there is no straightforward way for the WeightingInterface to be aware of the current epoch. While using trivial internal counters partially addresses the problem, they fail when batching is involved, as they would count each epoch multiple times.

Therefore, I propose adding the current epoch as an argument to the aggregate method of WeightingInterface, allowing it to be passed directly from the solver.

[dario-coscia]

We solved this in #620 by adding a link to the solver. Adding a link to the solver gives for free the possibility to access the trainer (solver.trainer) from which you can easily get the current epoch!