Multi GPU support for SAHI

[ArcPen]

Sahi is quite a good framework, I find it very useful to me. However, the lack of multi-gpu support limits its inference speed. I wrote a python script to make it able to run on multiple gpus. The main idea is to use symlink to divide the target image folder, and start an inference process for each gpu, then gather the results in json format, and convert into mmdet format.

I'm wondering if this method is needed, it seems kind of brutal but yet effective. If it is okay, I will try to intergrate it into the repo and make a pull request. Thanks.