Nash: Neural Adaptive Shrinkage for Structured High-Dimensional Regression