Clean up the embeddings API and parameter passing

Currently, there are keys around pretrained embeddings, projecting the embeddings, dropout, and so on, that are flat parameters to TextTrainer. There's also an embedding_dim parameter, which is a dict, with arbitrary allowed keys. We should make the flat parameters also a part of this dictionary, so the parameters look something like this:

"embeddings": {
  "words": {
    "dim": 100,
    "pretrained_file": "/path/to/glove",
    "fine_tune": false
  },
  "characters": {
    "dim": 16
  }
}

allenai / deep_qa

Clean up the embeddings API and parameter passing #307