Fixes on RecallAt metric for cut-off 1 and for scalar cut-off

[gabrielspmoreira]

Fixes #464, Fixes #700

Goals :soccer:

• This PR fixes RecallAt metric, that was giving wrong results for recall@1 (results higher than the upper cut-offs, and sometimes greater than 1) when 1 was among the cut-off values (e.g. RecallAt([1,5]).
• It also adds support to providing a scalar cut-off (e.g. RecallAt(5)), instead of a list or tuple.
• Finally, it fixes the error that happened when running trainer.evaluate() with a ranking metric with a single cut-off in the list (e.g.RecallAt([5]), as torchmetrics baseMetric` class we inherit from converts results with single element to scalar.

Testing Details :mag:

Added new unit tests to cover RecallAt metric, for both 2D and 3D inputs, with cut-off 1 and with scalar/unique cut-off. I have also used the paper reproducibility script to test it and now recall@1 is smaller than the upper cut-offs

[github-actions[bot]]

Documentation preview

https://nvidia-merlin.github.io/Transformers4Rec/review/pr-720