DNM: Detrend optimally combined data before running PCA

[tsalo]

Related to #1054.

Changes proposed in this pull request:

• Instead of just z-scoring the optimally combined data before running PCA, detrend it as well.

[tsalo]

• The three-echo PCA changes from 69 to 68 components, and the ICA (which converged before on the first seed) fails to converge in time for CircleCI.
• The four-echo test looks exactly the same, but that could be because it has GSR enabled.
• The five-echo PCA changes from 52 to 54 components, and the ICA (which converged before on the first seed) fails to converge in time for CircleCI.
• Logan's 8-echo checkerboard run (with aic+minimal) runs really well.
  • BTW running it with tedana_orig results in the clearly checkerboard-related component being rejected, which makes no sense to me. That said, that's true for detrending and no-detrending versions of the workflow.