MVN Normal example

[gcgibson]

It seems that derivative in the MVN normal example should be

return -1*np.matmul(theta-nm.repmat(self.mu, theta.shape[0], 1), np.linalg.inv(self.A))

which yields the estimate

ground truth:  [-0.6871  0.801 ]
svgd:  [-0.68918377  0.80174868]

Reference:

http://www.notenoughthoughts.net/posts/normal-log-likelihood-gradient.html