t-READi: Transformer-Powered Robust and Efficient Multimodal Inference for Autonomous Driving