
"Save individual results..." This option will compute one estimation of MSC for each epoch, using the data only for that epoch, using N windows (N depend on the window length and percentage of overlap). This option can be used to evaluate any change in coherence across epochs.

"Average crossspectra..." This option will compute one estimation of MSC for all the epochs, using the data from all the epochs, using N windows × numberofepochs, thus it is a better estimation of MSC.
The average of MSC results from Option1 are not the same as the MSC obtained with Option 2 as MSC is not linear.
The Coherence NxN [2021] processes uses the Fourier transform to compute the PSD and crossspectrum, to later compute MSC. Thus, MSC has the same resolution than FT. To obtain MSC in a band range, a different approach can be take, by filtering data and using Hilbert transform to compute the TF representation and compute the PSD and crossspectrum from that TF representation. More information on these differences here: Lagged coherence  #3 by Raymundo.Cassani
In a nottoofar update, the coherence process GUI will be updated to include both (Fourier transform and Hilbert (or wavelet) transform) approaches
Yes, with the process Frequency > Group in time or frequency bands