Compression methods

[xieyxclack]

This pr includes:

• Message compression based on grpc (gzip & deflate)
• Symmetric uniform model compression (32 bit -> 16/8 bit)
• Unittest for model compression in distributed & standalone mode
• Docs of usage of these compression methods

Thanks a lot for the help from @chenhesen!

[yxdyc]

LGTM. By the way, it is suggested to add (asymmetric) affine quantization in the next version, as it is a relatively simple and natural extension for the implemented symmetric uniform quant here, while can achieving smaller precision loss.