Torchmetrics usage improvements with classes instead of functionals

[DomInvivo]

Changelogs

• Relaxing the constraints on Torchmetrics version
• Changing the code to use the update and compute to avoid memory issues with large validation set
• Major changes and cleanup to the predictor_summaries.py to simply handle metrics updates and compute + new unit-tests
• Fixing all unit-tests, including those testing the training and finetuning of the models
• Adding more unit-tests to cover the changes in MetricWrapper
• Adding more unit-tests to cover SingleTaskSummary and MultiTaskSummary
• Adding more unittests in test_training.py and test_finetuning.py to make sure that the full pipeline runs
• Moving to a ProgressBar callback to fix the progress bar logging
• Running on a real task and ensuring that we recover the same performance, and that all metrics are reported correctly on CPU
• Running again on single GPU
• Running again with DDP multi-gpus, and that they give the expected values (all predictions are synced before computing the metrics)
• Update all the config files to be compatible with new dataloader, and some changes in torchmetrics (example, f1score and accuracy now require the task parameter.

---

Checklist:

• _Was this PR discussed in an issue? Issue Update MetricWrapper to work with the update/compute from torchmetrics #466
• Add tests to cover the fixed bug(s) or the new introduced feature(s) (if appropriate).
• Update the API documentation is a new function is added, or an existing one is deleted.
• Write concise and explanatory changelogs above.
• If possible, assign one of the following labels to the PR: feature, fix or test (or ask a maintainer to do it for you).

---

discussion related to that PR