Revisiting pre-trained remote sensing model benchmarks: resizing and normalization matters