Differentiate between binary and continuous outcomes about hete

[wlattner]

The easy path here is trusting the user-provided types. If y is a factor, the outcome is a binary (assuming only 2 levels), if y is numeric then the outcome is continuous, again assuming more than 2 distinct values.

y type	unique values	action
factor	2	-
factor	2+	error
numeric	2+	-
numeric	2	if 1/0 convert to factor + message

tmt type	unique values	action
factor	2	convert 1/0, use reference level for 0
numeric	2	verify only 1/0
numeric	2+	error
factor	2+	error

The user-supplied estimator must fit a classification model when y is a factor and a regression otherwise. Additionally, predict must return a vector. We can verify the predictions are in [0, 1] for the classification case.