CI : occasional multigpu test failure

[maleadt]

This happens once in a while on CI:

Some tests did not pass: 436 passed, 4 failed, 0 errored, 0 broken.
	base/array: Test Failed at /var/lib/buildkite-agent/builds/gpuci-16/julialang/cuda-dot-jl/test/base/array.jl:940
	  Expression: Array(a) == Array(b) == data
	   Evaluated: [0.0491718221481211 0.7675180540873912 … 0.8010551449181886 0.9725641606703476; 0.11907881640750706 0.08725304891274233 … 0.5641848519912193 0.1241890127157469; … ; 0.024094310524527707 0.8403841370820818 … 0.9772642076169137 0.9836615941923371; 0.6918572875342215 0.8907696748195567 … 0.09940363694051435 0.8025607099234905] == [0.0 0.0 … 0.0 0.0; 0.0 0.0 … 0.0 0.0; … ; 0.0 0.0 … 0.0 0.0; 0.0 0.0 … 0.0 0.0] == [0.0491718221481211 0.7675180540873912 … 0.8010551449181886 0.9725641606703476; 0.11907881640750706 0.08725304891274233 … 0.5641848519912193 0.1241890127157469; … ; 0.024094310524527707 0.8403841370820818 … 0.9772642076169137 0.9836615941923371; 0.6918572875342215 0.8907696748195567 … 0.09940363694051435 0.8025607099234905]

	base/array: Test Failed at /var/lib/buildkite-agent/builds/gpuci-16/julialang/cuda-dot-jl/test/base/array.jl:943
	  Expression: Array(a) == Array(b) == data
	   Evaluated: [0.0491718221481211 0.7675180540873912 … 0.8010551449181886 0.9725641606703476; 0.11907881640750706 0.08725304891274233 … 0.5641848519912193 0.1241890127157469; … ; 0.024094310524527707 0.8403841370820818 … 0.9772642076169137 0.9836615941923371; 0.6918572875342215 0.8907696748195567 … 0.09940363694051435 0.8025607099234905] == [0.0 0.0 … 0.0 0.0; 0.0 0.0 … 0.0 0.0; … ; 0.0 0.0 … 0.0 0.0; 0.0 0.0 … 0.0 0.0] == [0.0491718221481211 0.7675180540873912 … 0.8010551449181886 0.9725641606703476; 0.11907881640750706 0.08725304891274233 … 0.5641848519912193 0.1241890127157469; … ; 0.024094310524527707 0.8403841370820818 … 0.9772642076169137 0.9836615941923371; 0.6918572875342215 0.8907696748195567 … 0.09940363694051435 0.8025607099234905]

It's pretty fundamental that this would fail. I can't reproduce, even not on the CI server using exactly the same GPUs and CUDA toolkit...