Support for multi-output regressors

[janosh]

What's your stance on supporting multi-output regressors? Would enable use cases such as uncertainty estimation where the learner might predict both a mean and a variance.

This might not be that hard to implement. It would start with modifying this line in DFMLAdaptor.predict() and perhaps modifying the new output_col kwarg to also accept a list of strings, i.e. one name for each type of output the regressor generates.

[ardunn]

@janosh I'll have to think about this. My hesitation is that none of the current default configs support this action (i.e., will never produce multiple outputs) so the only people who'd be using it are those who are already making their own highly customized pipelines.

If it doesn't introduce too much code complexity and have one preset demonstrating this behavior, I'd be in favor of it. Since it's going to be a relatively large upgrade though, I'm marking this for the next major release.