Tensorflow-hub

* Enforce using Keras 2 by checking for installation of `tf_keras`
* Remove deprecated TF 1 API (`LatestModuleExporter`, `register_module_for_export`, ...)

* Require Python 3.9+.
* Bump rules_license to 0.0.4.
* Move `LICENSE` to `tensorflow_hub/LICENSE`.
* Use `license_files` instead of `license_file` in `setup.cfg`.
* Remove empty `tools` directory.

* Remove make_image_classifier and make_nearest_neighbour_index.
* Directly load https://hub.tensorflow.google.cn/* handles from GCS.

* Minor fixes for importing Keras and TF Estimator.
* Stop using distutils.version.
* Require protobuf >= 3.19.6.
* Require Python 3.7+.

* Improve support for `compute_output_shape` in `hub.KerasLayer`:
* This will now use the `output_shape` value provided on init, if present,
before falling back to the default behavior.
* Changes to make_image_classifier_tool:
* Add the option to read input with a tf.data.Dataset and use TF ops for
preprocessing. Enabling this feature requires TF 2.5 or higher.
* Set the default value of shear_range to 0. shear_range is deprecated and
incompatible when using TF ops for preprocessing.

* Use the Keras load context in keras_layer.py.
* Always use compressed model loading by default.
* Documentation improvements.