The default splice implementation is grossly inefficient and the iter based
ones work just fine, so use those instead.  I've measured an 8x speedup for
splice write (with len = 128k).
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>