Effnet docs release - #1074

This is wip to update the docs & release 2.1.10.

This is also benchmarked. My results were (Run with DLC 2.1.9 in Jan 2021) for 100 k iterations.

**ResNet50
Imgaug:
Results for 100000 training iterations: 95 1 train error: 2.13 pixels. Test error: 2.22 pixels.
With pcutoff of 0.4 train error: 2.13 pixels. Test error: 2.22 pixels

Scalecrop:
Results for 100000 training iterations: 95 2 train error: 1.47 pixels. Test error: 1.77 pixels.
With pcutoff of 0.4 train error: 1.47 pixels. Test error: 1.77 pixels

Tensorpack:
Results for 100000 training iterations: 95 3 train error: 2.09 pixels. Test error: 2.36 pixels.
With pcutoff of 0.4 train error: 2.09 pixels. Test error: 2.36 pixels

**EffNet-b3
Imgaug:
Results for 100000 training iterations: 95 4 train error: 2.39 pixels. Test error: 2.57 pixels.
With pcutoff of 0.4 train error: 2.39 pixels. Test error: 2.57 pixels

Scalecrop:
Results for 100000 training iterations: 95 5 train error: 2.26 pixels. Test error: 2.24 pixels.
With pcutoff of 0.4 train error: 2.26 pixels. Test error: 2.24 pixels

Tensorpack:
Results for 100000 training iterations: 95 6 train error: 1.65 pixels. Test error: 2.24 pixels.
With pcutoff of 0.4 train error: 1.65 pixels. Test error: 2.24 pixels

****NOTE: for the EfficientNet we have not optimized the learning rate schedule for batch size 1 and large images. Here we showed rigorously for different datasets and models that EffNets outperform ResNets and MobNets as expected from the imagenet performance: https://openaccess.thecvf.com/content/WACV2021/html/Mathis_Pretraining_Boosts_Out-of-Domain_Robustness_for_Pose_Estimation_WACV_2021_paper.html

(thereby we cross-validated the learning rate schedule for each model, you might want to do this as well on your data!).****