Physics-informed neural networks with residual/gradient-based adaptive sampling methods for solving partial differential equations with sharp solutions