Pure-Keras 3 conversions of Qwen2-MoE sparse Mixture-of-Experts LLMs (incl. Qwen1.5-MoE) (kerasformers).