BUG: CTF and proj (re)application

In principle we should be able to:

1. apply compensation or proj
2. compute forward on full set of channels
3. pick channels in data (maybe dropping compensation channels)
4. compute inverse without reapplying proj or compensation (#2310)

So we need to:

1. Always keep info['comps'].
2. Always keep info['projs'] complete with the set of channels that were used when they were applied, i.e., do not subselect channels if proj['active'].
3. If projs are applied to raw, for epochs(..., picks=[1], proj=True), raise an error telling people to use proj=False because they should not reapply?
4. Check for correspondence between comps and proj and data channels in the forward object during inverse application.

This non-reduction of the projs, however, would create some things to think about:

• In master, if you apply projs, then subselect channels, then compute the inverse, the projection operator is reapplied (I think?). This is a bit weird because it means that there were actually two spatial operators that got applied, but only the last (with fewer channels) gets applied to the lead fields.
• Currently we get the rank for inverse computation by subtracting the number of projectors, but this will only work if all channels are kept. We could still use n_proj -- we would just underestimate the rank, which is at least safer than overestimating it (and blowing up near-zero singular values).

This would probably take a lot of careful coding and testing to get right, but it should be doable. It seems like the cleanest option. @agramfort do you see any problems with this approach?