Timesformer reader align

[HydrogenSulfate]

1. add TimeSformer's performance in README.md and README_cn.md
2. add TimeSformer's config file in configs/recognition/timesformer/timesformer_k400_videos.yaml
3. add TimeSformer's en doc and Zh-CN doc
4. add TimeSformer's two image about main arch and attention arch
5. add UniformCrop in paddlevideo/loader/pipelines/augmentations.py, and corresponding UniformCropMetric in paddlevideo/metrics/uniform_crop_metric.py
6. add pyav's video decode method in paddlevideo/loader/pipelines/decode.py
7. add linspace_sample strategy in paddlevideo/loader/pipelines/sample.py
8. add VisionTransformer in paddlevideo/modeling/backbones/vit.py
9. update the different forward method in paddlevideo/modeling/framework/recognizers/recognizer2d.py while using TimeSformer and ViT based model
10. delete coloring and AverageMeter two unused import statement in paddlevideo/tasks/train.py
11. add an new training strategy gradient accumulation for TimeSformer and follow-up models, if used, in paddlevideo/tasks/train.py
12. fixed a bug when record the learning_rate in training process, when batch_size=1.
13. update the load_ckpt function for TimeSformer and follow-up ViT based models
14. update run.sh for TimeSformer
15. add a new InputSpec in tools/export_model.py for TimeSformer, even though there still remain some todo for inference model
16. add TimeSformer_Inference_helper class in tools/utils.py
17. add inference for timesformer and corresponding content in docs
18. fixed some hyper-link bugs in start.md
19. Update the description of TenCrop.
20. Change the order of data augmentation to align reader.
21. Add a new style of augmentation(almost in tensor mode), when backend is pyav
22. update utils.summary to fix some bugs where FLOPs could be negtive
23. Add trunc_normal_ function
24. Fix head/base.py line 108's bug

[huangjun12]

后面这三个参数用 **kargs，load接口里出现这三个参数有点奇怪

[huangjun12]

paddle.flops接口跑不通吗？