Entropy Adaptive Decoding: Dynamic Model Switching for Efficient Inference