The iterator counting assumed we're doing an obvious optimization when
only updating the refcount on indirect extents - but we're not doing it
yet.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>