Physics-Informed Machine Learning with Adaptive Grids for Optical Microrobot Depth Estimation