Introduce Eager-specific and Graph-specific Python TensorArray classes,

and refactor the Python TensorArray class to wrap an instance of one of
these two classes; the particular class to instantiate is chosen by
consulting whether the context is in graph or eager mode.

TensorArrays in Eager are simply Python lists of EagerTensor objects;
most operations on an _EagerTensorArray object pass through to array_ops.

This change is meant to ensure compatibility of Eager execution with existing
code bases (e.g., code using control_flow_ops) that use TensorArray objects. Eager users will be better served by maintaining their own lists of EagerTensors.

PiperOrigin-RevId: 173609453