Debiasing Classifiers by Amplifying Bias with Latent Diffusion and Large Language Models