FairPIVARA: Reducing and Assessing Biases in CLIP-Based Multimodal Models