A plain linear/relu stack ending in one logit, the same trunk
mlp_layers() builds for every other estimator in the package. The one
thing worth doing by hand is the initialization: scikit-learn draws both
weights and biases from Glorot uniform, torch uses Kaiming uniform on the
weight and a fan-in bound on the bias, and the fit has to start where
MLPClassifier's would for the numbers to line up.