Taming Many-Parameter BSM Models with Bayesian Neural Networks