Foundation Model-Architected Sparse Dictionaries for Fully Quantitative Interpretability in Whole Slide Image Decoding