Implement CPU->GPU events using `cuLaunchHostFunc`

Created by: vchuravy

Sadly this doesn't fix @lcw original issue:

    let kernel = happy(CUDA(), (1,))
        # do not precompile, this hangs since cuModuleLoadDataEx will block
        # if device is busy
        barrier = Base.Threads.Event()
        cpu_event = Event(wait, barrier)
        gpu_event = kernel(;ndrange=(1,), dependencies=cpu_event)
        notify(barrier)
        wait(gpu_event)
    end

we still hang in cuModuleLoadDataEx.