Seeing the Whole Elephant: Systematically Understanding and Uncovering Evaluation Biases in Automated Program Repair