Scalable Bayesian dynamic covariance modeling with variational Wishart and inverse Wishart processes