AtMan: Understanding Transformer Predictions Through Memory Efficient Attention Manipulation