This generates more realistic contents and avoids spurious HashCollisions when
generating a set of objects using these hypothesis strategies, at the cost of
slightly worse "boundary checking" (i.e. we won't check contents with a length >
4096 bytes).