`wait(kernel(...)` performs much worse on Julia v1.7

Created by: lmh91

Sometimes wait(kernel(...)) hangs up. MWE:

using KernelAbstractions
using CUDA, CUDAKernels
using ProgressMeter

@kernel function test_kernel(a::AbstractArray{T}) where {T}
    index = @index(Global)
    a[index] = T(1)
end

a = CuArray(zeros(Float32, 500, 500));
ndrange = size(a)

k = test_kernel(CUDADevice())
wait(k(a; ndrange))

@showprogress for i in 1:1000000 # get stuck at some point
    wait(k(a; ndrange))
end

I observed this on Julia v1.7.1 and v1.7.2. On Julia v1.6.5, however, it works and never gets stuck. Package versions:

(@v1.7) pkg> st
      Status `~/.julia/linux-ubuntu-20.04-x86_64/environments/v1.7/Project.toml`
  [052768ef] CUDA v3.8.3
  [72cfdca4] CUDAKernels v0.3.3
  [63c18a36] KernelAbstractions v0.7.2
  [92933f4c] ProgressMeter v1.7.1

I tested it also with different machines (Ubuntu and Windows) and different GPUs (all NVIDIA's).

As @maleadt suggested this might be related to https://github.com/JuliaGPU/CUDA.jl/issues/1350 and https://github.com/JuliaGPU/CUDA.jl/pull/1369.