Energy-Based Cross Attention for Bayesian Context Update in Text-to-Image Diffusion Models