Seeing the Invisible: Machine learning-Based QPI Kernel Extraction via Latent Alignment