Initially it was thought that we just wanted to ignore errors from
logged op replay, but it turns out we do need to catch -EROFS, or we'll
go into an infinite loop.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>