Learning non-Gaussian spatial distributions via Bayesian transport maps with parametric shrinkage