Switch to tf.keras [WIP]

[matthew-z]

DONE:

• switch to tf.keras and remove keras dependency
• upgrade ubuntu dist from 14.04 to 16.04 in CI, as tensorflow 2.0 requires at least 16.04
• upgrade setuptools in CI
• add keywords param in model.compile to allow users to disable some experimental tf2.0 features
• add pytest-xdist plugin to run each test in separate processes using fork (note that fork does not work in Windows)
• upgraded pytest to 4.4, as required by pytest-xdist

TODO:

• BiMPM and MultiPerspectiveLayer: need refactoring to avoid initializing layers in call, which causes errors with the tf 2.0 default eager mode.
• MatchLSTM: I think tf.layers.LSTM + model.fit/evaluate still have some bugs to fix, see https://github.com/tensorflow/tensorflow/issues/33258. Currently I set experimental_run_tf_function to false as a workaround. However, I still encountered some errors in CI from time to time (randomly happened), while local test is usually OK....

~PS: since there is no 2.3 dev branch, the current PR base is the master branch.~ changed the base to 2.3-dev

[bwanglzu]

@matthew-z I cleaned up all the archived branches, please send future pr to 2.3-dev.