during refbank seeding from top sb edge, blocks of width=4 are iterated twice

In the loop here, blocks (because they cross a SB boundary) are correctly indexed at 8x8 granularity, but the variable cand_mi_wide is in 4px units, not 8px. Because of that, if a block at top had a width of 2, it will be iterated twice. This may seem innocent, but the relevant called function [here](url) will update counters based on this. It would be slightly nicer if blocks of width 2 adjusted the index by 8px (2 mi units) so they wouldn't index themselves again. Is it still possible to suggest minor clean-ups like this?