ENH: Extending source-space-based reconstruction

After #8033, @agramfort had the idea to show how it can be used with data processed with ICA and also help with source space stats. So a couple of ideas:

1. Use it with ICA somehow (I guess we'd have to get the ICA "operator" matrix to be applied to the forward like a projection matrix?)
2. Add a way to reconstruct instances to do sensor-space stats on reconstructed data.

Maybe a general inst.reconsruct(info=None, operator=None) that allows you to remap to the same channels undoing the bias of projections (info is None, default) or to some channel positions (info is not None use case) assuming the given operator has been applied to the data (None means "get from info['projs'], non-None could for example be the ICA operator)? It would need to come with a warning in the docs that they should not be source localized, though, since the data will be rank deficient but not have anything in info['projs'] (except maybe EEG avg ref proj).

Start by reading issue #8033 and the existing reconstruction, projection, and ICA APIs. Clarify how an inst.reconsruct(info=None, operator=None) entry point would handle info['projs'], ICA operators, channel remapping, and rank-deficient outputs. Done should include both proposed use cases and the documented source-localization warning.