Update model loading to be compliant with new versions of tensorflow

[rmarquarlops]

This pull request includes :

1. Model loading update in order to be compliant with new versions of tensorflow :

• new models.py file which stores the model architectures;
• config.yaml modified accordingly (output bins are now necessary to load the neural models).

2. It is now possible to give a config_path in order to switch easily between model versions.
   ex: L2-wave-processor --input_path S1_...SAFE --save_directory /savedir --product_id E99 --config_path /localconfig_17km.yaml

3. Help display speed using L2-wave-processor -h or --help has been increased by managing imports.

[agrouaze]

The model path is optional, it means that, the processor is able to find the resolution of input L1B/L1C and to pick-up an default associated NN model?
What if a user try to predict wave parameters from L1B and NN model that don't have the same resolution (eg a 17.5km² L1B and a 5km² NN model ) ?

[rmarquarlops]

The model path is optional because the processor use the elements defined in the localconfig.yaml by default if anything is missing.
Nothing is done to prevent a prediction with a model which was trained on another resolution than a given SAFE.

To prevent that, it would be necessary to store somewhere which IDs (L1B, L1C) correspond to a model version.

[agrouaze]

I suggest to add a security assert to make sure the resolution of the tiles stored in the input file correspond to the model tile resolution. Do you think it could be done?
If the name of the model file already contains the resolution, it is may be enough.

[agrouaze]

We go for a additional informationgiving the version of the L1C product ID used to train the model used. This product-id will be printed at the begining of the script to perform predictions.