Residual-Quantile Adjustment for Adaptive Training of Physics-informed Neural Network