* update mae benchmark, we fintune the model four times, and chose the highest one as the result. we pretrain and fintune the model only by 8 V100 which is a little different from official one.
* readthedocs build support
* change opencv version from 4.4.0.40 to 4.1.2.30
* update config file path with github url
* fix: clone AutoAugmentPolicy from torchvision error
* reorganize requirements file
* add model compression tutorial and cls modelzoo to readme.md
* add optional req
* add mae large to mae benchmark
* fix dataset pipeline bug when timm using torhvision original pipeline in mae
* add mae fintune unittest
Co-authored-by: yanhaiqiang.yhq <yanhaiqiang.yhq@gitlab.alibaba-inc.com>