TODO for next refacto

• factorize SVR and autoencoder
• factorise Square template and Sphere
• latent vector added as bais (30% speedup)
• remove last th in decoder
• make large .pth tensor with all pointclouds (drop the nasty Chunk_reader)
• make-it multi-gpu (at least change all .cuda() to to.device()
• add netvision results
• rewrite main script object oriented like 3D-CODED
• check that everything works in latest pytorch version
• Add more layer by default and flag for the number of layers
• add latent vector as a bias for memory and time speed-up
• Add a flag to generate a mesh directly
• Add a python setup install ( that update the submodule, and install the right packages)