Nuisance regressors could get a second round of dimensionality reduction

[mwaskom]

It seems there is a lot of redundancy in the different sources of nuisance regressors added in #188.

For example, here is the cumulative variance explained when you have 6 components from each of the four separate sources, and then perform another PCA on the resulting matrix:

[mwaskom]

This suggests that most runs need ~15 (out of 24) components to explain 99% of the variance and ~10–13 to explain 95%.

[mwaskom]

If we implement this, it would be good to still plot the original nuisance data se we can understand where the different components are coming from.