Sometimes GraphemeCursor::prev_boundary returns PrevChunk but can't process anymore

[jmuk]

I got a sample code:

let s = "abcd";
let mut cursor = GraphemeCursor::new(3, s.len(), false);
assert_eq!(cursor.prev_boundary(&s[2..], 2), Err(GraphemeIncomplete::PrevChunk));
assert_eq!(cursor.prev_boundary(&s[..2], 0), Ok(Some(2)));

On the second assertion, it instead returns Err(GraphemeIncomplete::InvalidOffset). If it starts from position 4, as is written in the doc comment, it works. However it does not work when the cursor position starts from 3.