Convert SBD model to ONNX format [resolves #300]

[dafajon]

This PR consitutes of updates that relate to:

• Upgrading sklearn dependency and training a new .pickle model for sbd with the same input data.
• Serializing the new sklearn model with skl2onnx
• Loading the .onnx model and running inference with onnxruntime

sadedegel/bblock/doc.py

• ONNX runtime accepts FloatTensorType or Int64TensorType as value input for its DictVectorizer. Thus span_features types are all normalized to int.
• spans property is decoupled from need of generating sents. It is a cyclic dependency. This is also addressed with d6ab6aa.

sadedegel/bblock/util.py

• Implement a hasher util function for span features of type string.

sadedegel/ml/sbd.py

• Implement model saving with skl2onnx
• Implement model loading with onnxruntime
• onnxruntime session processes dict input one at a time instead of batches.

requirements.txt

• Add latest scikit-learn
• Add skl2onnx and onnxruntime