Sometimes a request fails because it is too large. The auto chunking function should prevent this from happening. It looks like it is not accounting for the mask byte when calculating bytes per pixel. I believe the fix is to add 1 to dtype_bytes to match how Earth Engine is calculating bytes per pixel. I'm currently testing a change.
Repro: https://gist.github.com/jdbcode/748f7bb5d861f5e096006c7dadbe10b4