Learning a Neural Solver for Parametric PDE to Enhance Physics-Informed Methods