The spec says with msaa disabled gl_SampleMaskIn is always 1.
This is not particularly related to arb_sample_shading, but drivers may
do different workarounds depending on the state of sample shading and
presence of bits in the shader which would force per-sample execution,
so it seems appropriate to test here.
At least r600/eg will fail (the hw will give the coverage mask for the,
pixel, not the coverage per sample or 1 if msaa is disabled).
Since the existing partition test could only be passed when giving the
wrong answer for this, I would expect more drivers to fail...

This also fixes the partition test with msaa disabled, as it expected a
wrong gl_SampleMaskIn value in this case (now omit the test in this
case, since it's quite likely multiple things are failing in
the test simultaneously and then the failures can't be distinguished -
the new test for this also doesn't require atomics).

Reviewed-by: Brian Paul <brianp@vmware.com>