Transform y into 2D tensor when y.ndim == 1 and likelihood == REGRESSION

[wiseodd]

Fixes #235

Rationale on why we don't want to throw an error:

• PyTorch works on both these 1D-output cases:
  • model(x).squeeze() with y.shape == model(x).shape[0:-1]
  • model(x) with y.shape == model(x).shape
• So, in the wild, users' dataloaders vary between those to y.shape's. It's too much to ask to require y.shape == model(x).shape since PyTorch doesn't.

[aleximmer]

Thanks for fixing it. However, I would slightly be in favor of either raising an error or at least a warning if out.ndim != y.ndim because that's the same behavior of MSELoss afaik?

Also a test is failing, any idea why?

[runame]

LGTM now