-Faster inverse with Cholesky decomposition.
-Bumped up default number of Fourier features for the conditioning set to 100.
-Added a permutation option to estimate the null (approx="perm"). May be better for smaller sample sizes, although the permutation doesn't change the fact that non-linear regression is difficult with few samples.
-Separated RCIT and RCoT into two different functions.
-Added KCIT from (Zhang et al., 2011).
Added an additional method for estimating the null distribution called "chi2" which normalizes the empirical partial cross-covariance matrix so that it asymptotically follows a Gaussian with diagonal covariance (after root-n multiplication). The resultant statistic therefore asymptotically obeys a central chi-squared distribution with 25 degrees of freedom rather than a weighted sum of chi-squares. I find that the p-values are slightly less accurate than the default method in general. On the other hand, the new method outputs a normalized statistic, so the statistics are comparable without the same random seed. This method may be useful in algorithms like HITON-PC which initially sort variables according to the test statistic.