Add way of skipping pretrained weights download

[epwalsh]

Fixes #4599.

Changes proposed in this pull request:

• Adds a load_weights: bool (default = True) parameter to cached_transformers.get() and all higher-level modules that call this function, such as PretrainedTransformerEmbedder and PretrainedTransformerMismatchedEmbedder. Setting this parameter to False will avoid downloading and loading pretrained transformer weights, so only the architecture is instantiated. So you can set the parameter to False via the overrides parameter when loading an AllenNLP model/predictor from an archive to avoid an unnecessary download.

For example, suppose your training config looks something like this:

{
  "model": {
    "type": "basic_classifier",
    "text_field_embedder": {
      "tokens": {
        "type": "pretrained_transformer",
        "model_name": "bert-base-cased",
        // ... other stuff ...
      }
    },
  },
  // ... other stuff ...
}

And now you have an archive from training this model: model.tar.gz. Then you can load the trained model into a predictor like so:

from allennlp.predictors import Predictor

overrides = {"model.text_field_embedder.tokens.load_weights": False}
predictor = Predictor.from_path("model.tar.gz", overrides=overrides)

[epwalsh]

Unfortunately this actually doesn't address #5170, because the SrlBert model uses the transformers library directly. But that's not hard to fix. I'll follow up with a separate PR for that in allennlp-models.

[ArjunSubramonian]

Awesome! looks great to me. thanks for meticulously going down the stack :)