Rate-In: Information-Driven Adaptive Dropout Rates for Improved Inference-Time Uncertainty Estimation