Update coco_evaluation.py

[congyao]

In the original script, the matching procedure between ground truth rectangles and detection rectangles is problematic. To be exact, in line 53 to line 66, in each iteration the function seeks the best match for each ground truth rectangle and removes the best match from the list of detection rectangles, if any. However, the one actually removed is NOT the best match, but the last one in the list of detection rectangles, because of this line: eval_bboxes.remove(eval_box_id). Here, match instead of eval_box_id should be used, since eval_box_id becomes the last index after the loop.