Add support for XavierLarningRateScaling in TF 1.14+

Optimizers tests are failing with TF 1.14.0rc0 and TF 1.14.0rc1.

Error output:

self = <larq.optimizers.XavierLearningRateScaling object at 0x133f54e80>, name = 'optimizer'
                                                                                                                                                             
    def __getattribute__(self, name):                                                                                                                        
      """Overridden to support hyperparameter access."""
      try:                            
>       return super(OptimizerV2, self).__getattribute__(name)
E       AttributeError: 'XavierLearningRateScaling' object has no attribute 'optimizer'
                                      
../venv-bop/lib/python3.7/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:532: AttributeError

[lgeiger]

If I recall correctly, I ran into some troubles since the learning rates in TensorFlow 1.14+ are now Tensors instead of floats, but I didn’t investigate further since TF 2 was still in alpha then. I pushed one (probably broken) workaround I had on my local system: https://github.com/plumerai/larq/blob/xavierv2/larq/optimizers.py

[lgeiger]

Beside implementing the new optimizer API, do you see any other difficulties to get this working with TF 1.14+? we have already updated the API for Bop optimizer and it shouldn’t be a huge hassle, IMO.

I also encountered errors due to these lines, but that might be because I was trying to support both v1 and v2 in one implementation. I can push my WIP branch from some time ago and we can take a look and see if there is a quick fix 👍