Describir: Enhancing Zero-Shot Learning Through Kernelized Visual Prototypes and Similarity Learning