The problem with the chosen depth values was that the depth values
naturally got viewport-transformed to values below 0.5 (for those
quads drawn) or above 0.5 (for those not drawn).
This allowed buggy implementations (in particular llvmpipe) to pass
the test, even though llvmpipe a) couldn't handle swapped near/far and
b) didn't actually do any clamping whatsoever in this particular case.
So change the depth values so that transformed values actually end up
"on the wrong side" of the depth range.

Tested-by: Ilia Mirkin <imirkin@alum.mit.edu>