TensorFlow 2 model fitting for negative binomial noise

[picciama]

Adds tensorflow 2 specific code at batchglm/train/tf2/.
Includes negative binomial noise model.
No changes in other folders were made.
To import, we will need to apply changes to api/models/glm_nb.py, so that it can import from batchglm/train/tf2/ instead of batchglm/train/tf/

[picciama]

Branch tf2_compatibility merged into tf2. Tf2 can now be merged with dev.
tf2 api has been added on the tf2_compatibility branch before, but doesn't affect tf1.

[davidsebfischer]

Checking locally:

git diff --name-status tf2..tf2_compatibility
D batchglm/train/tf2/init.py
D batchglm/train/tf2/base/init.py
D batchglm/train/tf2/base/estimator.py
D batchglm/train/tf2/base/external.py
D batchglm/train/tf2/base/model.py
D batchglm/train/tf2/base/optim.py
D batchglm/train/tf2/base_glm/README.md
D batchglm/train/tf2/base_glm/init.py
D batchglm/train/tf2/base_glm/estimator.py
D batchglm/train/tf2/base_glm/external.py
D batchglm/train/tf2/base_glm/layers.py
D batchglm/train/tf2/base_glm/layers_gradients.py
D batchglm/train/tf2/base_glm/model.py
D batchglm/train/tf2/base_glm/optim.py
D batchglm/train/tf2/base_glm/processModel.py
D batchglm/train/tf2/base_glm/training_strategies.py
D batchglm/train/tf2/base_glm/vars.py
D batchglm/train/tf2/glm_beta/init.py
D batchglm/train/tf2/glm_beta/estimator.py
D batchglm/train/tf2/glm_beta/external.py
D batchglm/train/tf2/glm_beta/layers.py
D batchglm/train/tf2/glm_beta/layers_gradients.py
D batchglm/train/tf2/glm_beta/model.py
D batchglm/train/tf2/glm_beta/processModel.py
D batchglm/train/tf2/glm_beta/vars.py
D batchglm/train/tf2/glm_nb/init.py
D batchglm/train/tf2/glm_nb/estimator.py
D batchglm/train/tf2/glm_nb/external.py
D batchglm/train/tf2/glm_nb/layers.py
D batchglm/train/tf2/glm_nb/layers_gradients.py
D batchglm/train/tf2/glm_nb/model.py
D batchglm/train/tf2/glm_nb/processModel.py
D batchglm/train/tf2/glm_nb/vars.py

All changes are in tf2, that looks fine.

[davidsebfischer]

waiting for #92

[picciama]

refactoring finished. The arguments for the training_method are not identical to the ones in train/tf1 which allows to access the method from within diffxpy.