External regressors with NaNs will crash decision tree

[tsalo]

Summary

I tried building a decision tree to use motion parameters from an fMRIPrep confounds file. The problem is that relative motion measures (e.g., framewise displacement) have a NaN in the first row of the confounds because they're derivatives. We could leave it up to users to deal with this, but I think an easy workaround on our end would be to backfill NaNs automatically.

Additional Detail

We could also raise a warning if NaNs are present to ensure users know what's happening.