LipGER: Visually-Conditioned Generative Error Correction for Robust Automatic Speech Recognition