PLS assumes # dims >= 25

The PLS neural regression benchmarks assume the model layer has at least 25 dimensions, and therefore fail to run when the model layer has fewer than 25.

Perhaps one could work around this issue by artificially padding the model layer with zeros, but imo it would be better to have this handled gracefully on the benchmark side by changing the sklearn constructor call to something like `PLSRegression(n_components=min(num_model_neurons, 25))`.

Not sure how to do this cleanly; submitting this issue as a suggestion.