Incorporate `add_embedding` into SummaryWriter

[Stonesjtu]

Some advantages to implement it in SummaryWriter:

• More general and uniform APIs to use.
• Manage all the embeddings under one dir: runs/XXX (currently the embedding must be written into separate directories.
• Don't have to run different tensorboard instances for embedding visualization and other summary info

And in your README.txt, the add_embedding is under SummaryWriter part.

If the original tensorboard don't have such mechanism, we can wrap the SummaryWriter and add_embedding into brand new unified APIs.

[lanpa]

The reason to separate embedding from other functions is that this function uses tensorflow's internal code to dump data, rather then SummaryWriter. (Because I fiddled around but in vain) And currently, the embedding folder depth can only be 1. This is a documented bug in a ongoing tutorial. XD. Another concern is pytorch/pytorch#2230, which breaks training procedure if you are using multi GPU, so I think it is still a premature porting. Finally, official tensorboard seems to unify data format as tensorSummary, if that comes true, we can solve the above problems at once then put it back to SummaryWriter class.