Upgrade Airline experiment to GPflow 2

[vdutor]

** DO NOT MERGE **

Hi all,
I spent a little bit of time updating the GPR_additive VFF model to GPflow 2 (2.0.4) and Tensorflow 2 (2.2.0) for the Airline experiment. IMPORTANT: I only updated the bits I needed for the Airline experiment, so this PR is very dirty and incomplete, but I still thought it could be useful for others. I would, however, not recommend merging this PR as it will leave the repo in a bad state.

I was able to exactly reproduce the MSE and NLPD given in the paper. The run times became significantly shorter. Experiment ran on a single 8GB GeForce GTX 1070 GPU and Intel i7-7700 CPU @ 3.60GHz CPU with 24 GB of RAM.

	N	10,000	100,000	1,000,000	5,929,413
MSE
	mean	0.89337489	0.81916289	0.83120151	0.82781694
	std	0.15461227	0.05040159	0.01323845	0.00437594
NLPD
	mean	1.36232179	1.31913063	1.32644348	1.3244345
	std	0.0906293	0.03023875	0.00791944	0.00265458
Timing (clock)
	mean	18.3634206	23.1457316	58.2810024	277.9453202
	std	2.79274855	2.39547268	2.23852881	2.93161274
Timing (time)
	mean	6.78150756	7.92476165	17.44419279	75.61273456
	std	0.85582754	0.4776275	0.4343648	0.75147633]