[Impeller] Failing GPU compute tests on Intel macOS

[zanderso]

@dnfield

[ RUN      ] Compute/ComputeTest.HeartCubicsToStrokeVertices/Metal
../../flutter/impeller/renderer/compute_subgroup_unittests.cc:163: Failure
Expected equality of these values:
  q->count
    Which is: 0
  golden_heart_quads.size()
    Which is: 26
../../flutter/impeller/renderer/compute_subgroup_unittests.cc:177: Failure
Expected equality of these values:
  p->count
    Which is: 0
  golden_heart_points.size()
    Which is: 147
../../flutter/impeller/renderer/compute_subgroup_unittests.cc:187: Failure
Expected equality of these values:
  v_count->count
    Which is: 0
  golden_heart_vertices.size()
    Which is: 586
[  FAILED  ] Compute/ComputeTest.HeartCubicsToStrokeVertices/Metal, where GetParam() = 4-byte object <00-00 00-00> (322 ms)
[ RUN      ] Compute/ComputeTest.QuadsToPolyline/Metal
../../flutter/impeller/renderer/compute_subgroup_unittests.cc:317: Failure
Expected equality of these values:
  p->count
    Which is: 0
  golden_heart_points.size()
    Which is: 147
[  FAILED  ] Compute/ComputeTest.QuadsToPolyline/Metal, where GetParam() = 4-byte object <00-00 00-00> (8 ms)

[dnfield]

I'm going to try to get a hold of an Intel Mac mini to test this locally on. Otherwise I may have to SSH into one of the bots after quarantining it.

Our oldest bots are on Intel Iris 5100 GPUs, which should be enough to work with these tests.

[dnfield]

I am in position of the machine and waiting for some updates to be able to fully provision it. Hopefully I can reproduce things there once that's done.

[dnfield]

There are (at least?) three bugs:

1. The logic that's guarding the thread pool size is doing truncating division without guarding against truncation to zero. Easy fix.
2. The shaders are abusing barrier. I didn't think I was following the spec but it just worked on the M1 laptops so I figured maybe I misunderstood the spec. I was, in fact, misundersatnding the spec and getting lucky.
3. On the Intel IRIS 5100, the threadExecutionWidth is getting cut down to half of what the M1 uses for the same shader, which means the subgroup isn't big enough to handle all the math in one go. I need to fix that.