We should remove "outputs" from the problem description

It is never really used in the core package runtime or other code, or even in the rest of the program. It is something which was defined in MIT-LL problem description format, but I have not seen it really used there really. Where things are stored is generally dependent on how pipeline is run and not on the problem. We do not use that in TA2-TA3 API, so this also makes problem description send over the TA2-TA3 API different from one loaded from the disk. (See https://gitlab.com/datadrivendiscovery/ta3ta2-api/issues/106.) Also, I do not think it will generalize well if we will have loaders from other sources, like OpenML, which do not have such a concept. Moreover, I do not think such information is really necessary to be stored in metalearning database.

So I would suggest we just remove it. Similarly to how we are not preserving data splits information in problem description (while it is in MIT-LL problem description format) but require one to just pass that information through other arguments when running runtime. I see similarly information about where to sore outputs.

If anyone is using outputs in problem description, please say so.