ENH/MAINT: CNN regionalization, smash.Net.add() is deprecated

[nghi-truyen]

• New regionalization mapping with CNN
• smash.Net.add() is deprecated and replaced by multiple addition functions: add_conv2d, add_dense, add_scale,... The activation layer is now added by passing an argument into a trainable layer.
• New unit test for smash.Net() combining Dense and Conv2D layers.

Important notes:

• The baseline was changed for all optimization functions with ANN due to the change of the shape of neural network weights (W.T instead of W). This modification is intentionally made for the consistency with the weights of the neural network that would be added into the forward hydrological model in view of ENH: New hybrid and ODE-based hydrological models for gr4 structure #253.
• The tests for ensuring the precision of the feed forward and backward pass of the neural networks written in smash are on this git:
  https://github.com/nghi-truyen/test-smash-net

[nghi-truyen]

@inoelloc Hello,
Je viens de faire 2 PRs sur le changement de l'optim Fortran-Python et l'ajout de CNN (donc la méthode net.add a aussi changé). L'idée c'est de réduire au maximum les conflits sur ma branche dev (modèle hydride) et de faciliter les changments à faire pour chaque PR. Je propose de faire les PRs par cet ordre pour simplifier le boulot :

• Moving L-BFGS and SBS optimizers from Fortran to Python #238
• Add CNN mapping for regionalization #251
• ENH: New hybrid and ODE-based hydrological models for gr4 structure #253
• Combine all optimizers from Fortran and Python #240
• mapping="ann" would be deprecated and replaced by "mlp" and "cnn" #254
• Add a callback signature into the optimization methods to get intermediate results #252

[inoelloc]

@inoelloc Hello, Je viens de faire 2 PRs sur le changement de l'optim Fortran-Python et l'ajout de CNN (donc la méthode net.add a aussi changé). L'idée c'est de réduire au maximum les conflits sur ma branche dev (modèle hydride) et de faciliter les changments à faire pour chaque PR. Je propose de faire les PRs par cet ordre pour simplifier le boulot :

* [Moving L-BFGS and SBS optimizers from Fortran to Python #238](https://github.com/DassHydro/smash/issues/238)

* [Add CNN mapping for regionalization #251](https://github.com/DassHydro/smash/issues/251)

* [New hydrid and ODE-based hydrological models #253](https://github.com/DassHydro/smash/issues/253)

* [Combine all optimizers from Fortran and Python #240](https://github.com/DassHydro/smash/issues/240)

* [mapping="ann" would be deprecated and replaced by "mlp" and "cnn" #254](https://github.com/DassHydro/smash/issues/254)

* [Saving the Model object and model ddt through the optimization process #252](https://github.com/DassHydro/smash/issues/252)

Hi, thanks. ok ok pour tout ca. J'aimerais finaliser les PR en cours avant de merge tout ca ^^"

[inoelloc]

Voila pas grand chose finalement. Je suis quand meme partant pour qu'en discute un peu ;)