[Feature]: Refactor Predictor subclasses into parameterizations via OutputTransforms

[hwpang]

Notes
From @davidegraff 's comment from #726

2. Expressing our Predictor subclasses as parameterizations of a base Predictor using different OutputTransforms rather than separate subclasses. For instance, a BinaryClassificationFFN is functionally the same as a RegressionFFN with the only differences of:

   • use of a BCE loss function
   • use of a sigmoid activation activation at test-time rather than an inverse normalization (i.e., a SigmoidTransform)