Improved self-calibrated signal model for χ-separation using single-subject (N=1) physics-constrained deep learning with histological validation