Split Yolov5 in GPU/CPU versions

While working on GPU support for a couple of executors in this [issue](https://github.com/jina-ai/executors/issues/169), I've stumbled over a problem for the [Yolov5 segmenter](https://github.com/jina-ai/executors/blob/main/jinahub/segmenters/YoloV5Segmenter/yolov5_segmenter.py), which can be seem in this [draft pr](https://github.com/pulls).
CI always fails to install pycocotools on our custom [GPU github action runners](https://github.com/jina-ai/jina-terraform/tree/main/gpu-github-runners).
That is fairly annoying, because we dont even need this dependency, it is only introduced because the original upstream [authors of yolov5](https://github.com/ultralytics/yolov5) dont provide a Pypi package and we are using this [fork](https://github.com/fcakyon/yolov5-pip) instead. There also have been some related discussions already: [here](https://github.com/ultralytics/yolov5/issues/2525) and [here](https://github.com/ultralytics/yolov5/issues/2481)

So this issue contains two sub tasks:

- [ ] Ideally submit a PR to the original yolov5 repo and provide a setup.py
- [ ] Replace the yolov5 fork with the real one and do necessary changes for separing gpu/cpu version of the yolov5 segmenter