feat: tensorflow wrapper

[yimuchen]

Adding a wrapper for running tensorflow inference models

• [] Properly handle treatment of missing values. Tensorflow does not work with missing values. User defined code will require explicating padding with dummy values.
• Dask awkward type tracing generates mismatched dimensions. User side definitions will need to define ensure the arrays passed will have the correct dimensions.

Discussion of tensorflow operations:

• Add accelerator device handling (do we need to?)
• Moving to use model.predict instead of model.__call__ [1]. The current implementation has issues with model.predict, in particular with type tracing arrays. Do we need to do this if batching is already handled by dask anyway?

[1] https://keras.io/getting_started/faq/#whats-the-difference-between-model-methods-predict-and-call

[yimuchen]

Added documentation for common data mangling patterns.