Improve auto-correlation estimation during regression

Currently, we do a pretty crude ARMA model if the user wants to model auto-correlations. We have a few options to improve this:

1. AFNI style: implement our own limited grid search of ARMA parameters per voxel rather than unconstrained optimization in statsmodels (currrent) - probably the trickiest and slowest
2. Implement pre-whitening routines by first estimating OLS, then using the residuals of order N to adjust the data, and finally fit OLS again. Nistats has an implentation of this: see this and this.
3. Port over SPMs new FAST routine which does something like using basis-sets of covariance matrices to account for auto-correlation