BayesianLinear prediction fixes and enhancements
BayesianLinear prediction was pretty screwed up: the distribution of predicted values was messed up, and error variances weren't being incorporated at all. This fixes it. Since this is a pretty major fix, bumping the eris version. This also changes the predict interface to be able to predict multiple rows at once (instead of just one row), and adds a new predictGeneric() method that allows prediction of arbitrary functions of y*. One such implementation -- which returns both the mean and variance of the prediction values -- is included in the new predictVariance() method. This, however, breaks API compatibility, so bumping the lib version. linear-draw-test is a new scratch script that verifies that the distributions are actually what Koop says they should be. This also adds a (non-truncated) multivariate t generator to go along with multivariateNormal(), and a rmvt scratch script to test it.
Showing with 423 additions and 73 deletions