Add pretrained models of waifu2x

[Witiko]

We are currently shipping with only the ESPCN, SRCNN, VESPCN, and VSRNET models, which are pre-trained on real-world footage. Where superresolution really shines, however, is text, CGI, and drawings. Waifu2x is a good early implementation of superresolution and its pretrained models for drawn content are the industry standard. Although waifu2x uses the Torch Lua framework, a Python reimplementation in Tensorflow exists and supports loading the pre-trained models.

We should:

• write Python code that would load the pretrained models of waifu2x and convert them to the format expected by libtensorflow,
• download and convert the pretrained models of waifu2x during script/produce-sr-models, and
• add visual comparisons and code examples for the different models to the readme.

[Witiko]

See also tcyrus/waifu2x-onnx, which documents the conversion of pre-trained waifu2x models to the ONNX format.