to/from UTF-8, instead of the locale encoding to avoid inconsistencies with
os.fsencode() and os.fsdecode() which are already using UTF-8.