[ENH] Implement pure-Python ICA with mixture modeling

[tsalo]

Closes #2.

TODO:

• Incorporate license and attribution info for nipy's GGM.
• Convert pseudocode to real code (this mostly means supporting images in addition to arrays and getting computefeats2 working).
• Compare mixture modeling results to MELODIC.
• Add tests.
• Integrate changes into workflow.

Changes proposed in this pull request:

• Add GGM class from nipy for mixture modeling.
• Use tedana for PCA + ICA.

[tsalo]

More aggressive masking seems to be necessary. I noticed that the ICA components were showing up outside the brain with the old method.

[eurunuela]

As discussed with @CesarCaballeroGaudes yesterday, the mixture modeling is not necessary but it helps with the accuracy.

The ICA-AROMA paper states:

To increase accuracy of the spatial features, IC spatial maps are first thresholded using a Gamma/Gaussian alternative testing approach (p N 0.5). This approach is automated in FSL's MELODIC.

So, we could try running the workflow without the mixture modeling first and see if the results look good. If we see the mixture modeling is critical for accuracy, we can open a new PR with it.

[tsalo]

To be honest, I'm not confident that either the features or the classifier in AROMA are robust enough to handle inputs that don't exactly match the original method. From a conceptual standpoint, unthresholded maps may be fine, but from a practical one, I'm not so sure.

[tsalo]

I have ME-ICA/aroma#22 open now, so I'm going to close this.