perf: parallel folder walk

Workers now read folders as well as format files, from one shared
queue, so formatting starts before the walk ends. Only changed and
failed files are kept and sorted, and the list is written at once
instead of one write per line.

On 6,784 XML files in 4,203 folders, all cores: 205 to 166 ms.
The HTML corpus: 6.8 to 4.8 ms.