PAGP: A physics-assisted Gaussian process framework with active learning for forward and inverse problems of partial differential equations