Added support for masked language modeling (bidirectional models)

[shehadak]

This PR builds on the Huggingface subject, which assumes that models are autoregressive (following the ModelForCausalLM). This PR adds support for bidirectional models with masked language modeling following the ModelForMaskedLM). Since bidirectional models rely on future context, I use a sliding window approach (see google-research/bert#66). In particular, for each text part, up to w/2 tokens are included for the current part + previous context, and the remaining w/2 tokens are masked.

The region_layer_mapping for the language system was determined by scoring every transformer layer in BERT's encoder against the Pereira2018.243sentences-linear, Pereira2018.384sentences-linear, and Blank2014-linear benchmarks, and choosing the layer with the highest average score.

This PR also provides unit tests for reading time estimation, next word prediction, and neural recording, using the bert-base-uncased model. Future models can use the same format, as long as they implement the ModelForMaskedLM interface. For example, to add the base DistilBERT model:

model_registry['distilbert-base-uncased'] = lambda: HuggingfaceSubject(model_id='distilbert-base-uncased', region_layer_mapping={
    ArtificialSubject.RecordingTarget.language_system: 'distilbert.transformer.layer.5'}, bidirectional=True)