Exploring the Design Space of Machine Learning Models for Quantum Chemistry with a Fully Differentiable Framework