Update spaCy to use the revamped thinc 8.0.0 currently on thinc's develop branch, see explosion/thinc#143

Description

• allow training from config, cf example configurations in https://github.com/explosion/spaCy/tree/feature/config/examples/experiments/ptb-joint-pos-dep
• Tok2Vec Pipe component and Tok2VecListener Model
• update layer names according to the thinc refactor, e.g.
  • Pooling(sum_pool) → SumPool
  • with_flatten → with_array
  • with_square_sequences → with_list2padded
  • flatten_add_lengths → list2ragged
  • Linear → SparseLinear
  • Affine → Linear
  • ...
• dimensions of models (e.g. nI and nO) are queried through model.has_dim(), model.get_dim(), model.set_dim(). Similar functions exist for _param (e.g. weights), _attr (e.g. window_size - the data has to be json-serializable) and _ref (e.g. a model's "tok2vec" component).
• quick reimplementations of the previous functions in _ml can now be found in component_models
• dropout is defined much more consistently across different layers, ultimately translating in a Dropout layer with a certain rate that can be set across the model by set_dropout_rate(model, drop)
• use ml_datasets for loading example datasets
• added "overfitting" tests for parser, tagger, ner etc, to check whether they can quickly converge on just a handful of training examples (taken from the docs)

Types of change

enhancement / refactor

Tasks

Some are done, some ongoing, some TODO

• Fix all tests (normal and slow)
• Run train_from_config with defaults.cfg : runs fine
• Run train_from_config with various other configs
• Ensure EL algorithm runs & learns
• Ensure tagger runs & learns
• Ensure parser runs & learns
• Run train script with new branch
• Run pretrain script with new branch
• Run other CLI scripts
• Clean up example configs in code

Checklist

• I have submitted the spaCy Contributor Agreement.
• I ran the tests, and all new and existing tests passed.
• My changes don't require a change to the documentation, or if they do, I've added all required information.