Analysing the Residual Stream of Language Models Under Knowledge Conflicts