Scope `BertSpanLabeler` and supporting infra

[jbischof]

BertClassifier supports one output for the entire input sequence, but span labeling is an important research benchmark. This will require some scoping and design:

1. Do we need a separate BertSpanLabelingPreprocessor or just a more complicated label to pass through?
2. Do we want to offer a script that can turn a standard dataset like squadv2 into something our preprocessor can use? The "answer text" must be turned into indicies in the tokenized input.
3. Create a BertSpanLabeler task model with the correct task-specific layers and preprocessing.

[jbischof]

References:

• tensorflow/models has BertSpanLabeler: https://www.tensorflow.org/tfmodels/nlp#span_labeling_model
• HF has BertForQuestionAnswering: https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertForQuestionAnswering