SpeechBrain should rely on steps and epochs

Describe the bug

As discussed during the SB meeting, SpeechBrain has been built around the concept of epoch, which now is starting to be outdated as many datasets are now too large to be run on a full epoch (e.g. libriheavy, you might prefer doing 0.5 epochs).

I propose to investigate this issue by enabling speechbrain backend to either rely on epoch or step (or both depending on the context). For instance, we should be able to specify a specific number of steps for training (and evaluation).

Expected behaviour

.

To Reproduce

No response

Environment Details

No response

Relevant Log Output

Additional Context

No response