I agree that the reason that the TFR.plot and TFR.plot_topo functions have a baseline and mode params is to emphasize the change in power between baseline period and the rest of the trial.
@mscheltienne is correct that you could use TFR.apply_baseline() to do that step before doing statistics (provided that the correct baseline time span is still present, i.e., you haven’t cropped or shifted the epochs).
As to whether you should do this before clustering — I think generally it’s probably the right thing to do (e.g., we do it in this example: Non-parametric 1 sample cluster statistic on single trial power — MNE 1.6.1 documentation), because generally what neuroscientests are interested in is something like “increase in power over baseline in frequency band X being greater in condition 1 than in condition 2”. In other words, the actual value of the power estimate is usually not what we care about, rather it’s usually the relative power in different bands, conditions, or subject populations.
I think @agramfort has some strong opinions about which method of TFR baseline correction are appropriate, so if you have follow-up questions about that hopefully he can chime in.