*   Add TensorSpec to represent the specification of Tensors.
    *   Constant folding pass is now deterministic.
    *   Support `float16` `dtype` in `tf.linalg.*`.
    *   Add `tf.estimator.export.TensorServingInputReceiver` that allows
        `tf.estimator.Estimator.export_savedmodel` to pass raw tensors to model
        functions.

## Deprecations

This operation does not have an associated kernel and is not intended to be
executed in a regular TensorFlow session. Instead it is intended to be used for
testing or for special case where a user intends to pass custom MLIR computation
through a TensorFlow graph with the intent of having custom tooling processing
it downstream (when targeting a different environment, like TensorFlow lite for
example).