Feature: TL1to2 migration

[tiancheng2000]

Migration points:

• keep and comment out important TL1 related codes, so can also be used as migration tutorial.
• use dynamic model and auto-graph(@tf.function) train step. (key points for TL2 migration)
• directly reuse VGG19 from tensorlayer\models\vgg19.py instead of handmade.
• use DevConv2d instead of Conv2d as decoder (optional).
• use save_weights/load_weights instead of tl.files.save_npz/load_npz.
• not migrated:
  • replace inputs feeding with tf.data.Database
• other helpful stuffs:
  • a pretrained vgg16 weights h5 (13.4M) to avoid download problems
  • pytest codes
  • auto-save on keyboard interruption

[tiancheng2000]

Hi, original author and reviewers, I re-run my TL2 version recently, and found that even after a 6-hour model training, the weights was already good enough to produce some impressive outputs (as show below). HTH!

[zsdonghao]

@tiancheng2000 you can merge the PR, if you are sure others can run the code.