Acknowledge CUDA_DEVICE_QUERY during GPU selection

[denisalevi]

Before, CUDA_DEVICE_QUERY had not effect for GPU selection in brian2cuda. This PR should fix that.

This needs some tinkering around with the different IDs (the ones reported by nvidia-smi, which ignores CUDA_VISIBLE_DEVICES) and the ones defined by CUDA_VISIBLE_DEVICES (which e.g. deviceQuery takes into account).

Here is the updated doc for the gpu_id preferences. This is how it should work. What is missing is that with prefs....gpu_id = None and CUDA_VISIBLE_DEVICES set, it still ignored CUDA_VISIBLE_DEVICES currently.

    The ID of the GPU that should be used for code execution. Default value is
    `None`, in which case the GPU with the highest compute capability and lowest ID
    is used.

    If this preference is set, it has to be the ID reported by `nvidia-smi`, which
    ignores the environment variable `CUDA_VISIBLE_DEVICES`.

    If this preference isn't set, `CUDA_VISIBLE_DEVICES` is not ignored. E.g. with
    `CUDA_DEVICE_QUERY=1,2` only GPUs 1 and 2 will be considered during GPU
    detection.

[denisalevi]

I thought cudaSetDevice would ignore CUDA_VISIBLE_DEVICES. But it doesn't. Hence the most obvious and easiest solution here is to never ignore CUDA_VISIBLE_DEVICES. This is implemented now. The only place where this was failing was when detecting all available GPUs using nvidia-smi, which ignores CUDA_VISIBLE_DEVICES. Now, detection of all devices checks CUDA_VISIBLE_DEVICES as well.