Human-in-the-Loop Optimization for Deep Stimulus Encoding in Visual Prostheses