Handle discrete labels correctly & switch to GridSearchCV for readability

[anandawolz]

• Leverage KNNDecoder’s built-in classifier/regressor selection via dtype casting
• Simplify continuous vs. binary detection and metric computation
• Replace manual k-loop and 8/9-split with a single GridSearchCV

when plotting the results using built-in functions of CEBRA-lens the axis are not adapted yet, and still show R^2 score, also for classification, but the result is correct now --> will be for a later change

[CeliaBenquet]

This needs to pass the tests first (that's for that PR I wanted to add them :))

[CeliaBenquet]

As discussed with @anandawolz, those changes were transferred and improved in #59. Can be closed.