Intermittent GC-related test failure (isempty(cuda_ptrs))

[vchuravy]

@maleadt noticed that every so often the CI would fail due to a test error in gc.jl.
The specific test that fails is: https://github.com/JuliaGPU/CUDArt.jl/blob/7c7157019fa69539b2547b5317d9770c89d7d462/test/gc.jl#L57

g = CUDArt.CudaArray(a)
# In case the next gc() test yields an error, the next two lines let
# us do some archaeology
dictcopy = deepcopy(CUDArt.cuda_ptrs)
gptrcopy = copy(pointer(g))
gc()   # Check that this doesn't delete the new g
@test !isempty(CUDArt.cuda_ptrs)
h_g = CUDArt.to_host(g)
CUDArt.free(g)

With a little bit of debug information added and TRACE=1 I get:
while TRACE=1 julia --compilecache=no -e 'Pkg.test("CUDArt")'; do :; done

TRACE: Malloc CudaPtr Ptr{Void} @0x0000002303ee0000
TRACE: Finalizing CuContext at Ptr{Void} @0x00007f06843d85d0
TRACE: Not destroying context CUDAdrv.CuContext(Ptr{Void} @0x0000000003f03500,false,false) because we don't own it
TRACE: Freeing CudaPtr CUDArt.CudaPtr{Float64}(Ptr{Float64} @0x0000002303ee0000,CUDAdrv.CuContext(Ptr{Void} @0x0000000003f03500,false,false))

I am currently travelling but should get to this before long.

[maleadt]

Did some testing, this bug is also present on 0.2.3 + Julia 5. So even though it would be nice to fix it before release, we're blocking way too many other packages.

[maleadt]

Seeing how this is still broken, I'm going to disable the test since it makes it much harder to judge PRs etc.