On Large-Scale Dynamic Topic Modeling with Nonnegative CP Tensor Decomposition