@print functionality inconsistent with CUDADevice()

This is more for my own reference later.

If I create the following kernel:

@kernel function check_kernel()

    cI = @index(Global)

    if cI == 2
        @print("cI is: ", cI, '\n')
    end
end

function check(use_gpu)
    if use_gpu
        kernel! = check_kernel(CUDADevice(), 256)
    else
        kernel! = check_kernel(CPU(),8)
    end
    kernel!(ndrange=256) 
end

I find odd behavior in the REPL:

julia> include("src/check.jl")
cI is: 2KernelAbstractions
.CudaEvent(CuEvent(Ptr{Nothing} @0x000055d1cee5f4c0, CuContext(Ptr{Nothing} @0x000055d1ce966110)))

julia> wait(check(true))

julia> wait(check(false))
cI is: 2

julia> exit()
cI is: 2

For some reason, the printing for kernels run on a CUDADevice() will only print on exiting the REPL

I guess there is some way to make the wait() function also prompt the @print macro, but it might take some digging.

Reproduce the behavior from the example in src/check.jl, comparing @print output for CUDADevice() and CPU() around wait(check(true)) and wait(check(false)). Read the @print and wait paths first; done means CUDA kernel output is available after the corresponding wait rather than only when the REPL exits.