Estimates on the generalization error of Physics Informed Neural Networks (PINNs) for approximating PDEs