[ENH] Adding STC classes that support frequency and trial dimensions

Talking to @britta-wstnr about the options for time-frequency resolved source estimates (and also trial-resolved source estimation ideally for forward extensibility) here is what we came up with:

Here is the problem:

• If you want to resolve source estimates over frequencies as well as time (and trials), you get a list of stcs (or even a list of lists of stcs) which contain a lot of redundant information about the forward operator etc. and are a mess to check that the stc attributes all match every time
• Adding dimensions to the stc data object would be the easiest but would break backward compatibility for people using the public attribute stc.data

Here is the proposed solution:

• Make three new classes for surface, volume and mixed source estimates that have a dimension for frequencies and trials (could be just a flat dimension if it's not used) and show how to assemble from beamformers in an example

Avoiding potential issues:

• Returning a 4D stc object directly from the beamformer code probably is not a good idea as it is likely not feasible in RAM for most setups to compute it all at once so it probably makes more sense to allow the user to assemble the data in a way that most efficiently trades off between memory use and speed

We couldn't think of any other future extension dimensions that would be good to include since this proposal is to make new classes but if there are things, we could consider including them. Connectivity comes to mind but I think should be outside the main MNE-Python repository. TFR seems right in the MNE-Python main repository wheelhouse so I think this would be nice to incorporate so that the visualization GUI can take in an object with all the necessary info and the user doesn't have to cobble it together. What do people think?

cc @agramfort @larsoner @jasmainak related to GSoC

Start by locating the existing surface, volume, and mixed source-estimate classes and the beamformer assembly paths. Review how current STC data and metadata are represented, then define the required frequency and trial dimensions and an assembly example. Done means the three proposed classes and documented assembly approach support these dimensions without changing the existing public stc.data behavior.