Sparse Bayesian long short-term memory networks for computationally efficient stochastic modeling of plasma etch processes