BUG: Only add decimate to correct classes

[larsoner]

So it turns out the raw.decimate was unintentionally added by @alexrockhill in #10945 when refactoring TimeMixin. When I went to test it, it didn't work -- it relied on presence of a ._raw_times attribute which doesn't even exist for Raw instances. Same thing with SourceEstimate classes, which also errantly got .decimate methods -- they didn't work and I don't think were intended to be added. Since they never worked for these other classes, we can remove them without a deprecation cycle.

A future option would be to add proper support for them -- I actually did recently try to use raw.decimate because I had suitably filtered my data but ended up needing to use the much slower .resample instead! -- but it's not trivial. We jump through some substantial hoops in raw.resample to deal with concatenated instances, so not worth it for now.

Thus in this PR I:

1. Put TimeMixin back to what it was, where the only public method was time_as_index
2. Added all the stuff from [MAINT, MRG] Refactor TimeMixin to include more time operations, ignore filter warning for decim for EpochsTFR #10945 in a new class ExtendedTimeMixin(TimeMixin):
3. Used the ExtendedTimeMixin as the mixin class for Epochs, Evoked, EpochsTFR, EvokedTFR, EDIT: and DipoleFixed. This means .decimate will only exist for those, which I think was the intended effect.
4. Fixed a bug where .first was errantly being set on Epochs (this is an Evoked-only attribute)
5. Properly tested epochs and evoked decim
6. Properly tested AverageTFR decim (which was broken on main)
7. Properly tested *TFR.shift_time which was available but untested on main

Closes #11797

[agramfort]

Thx @larsoner 🙏

[alexrockhill]

Thanks @larsoner and my bad for not thinking to check that TimeMixin was used in Raw. Thanks for fixing the bugs.