Modularizing encoder/decoder in Autoencoder class

[dreamer2368]

This PR must be merged after PR #9 .

Using DistributedDataParallel for data parallelism requires access to encoder and decoder as a torch.nn.Module. Current Autoencoder class provide encoder and decoder as member functions, though DistributedDataParallel cannot use custom member functions except forward.

• lasdi.latent_space.MultiLayerPerceptron is now provided as a distinct module for a vanilla MLP.
• lasdi.latent_space.Autoencoder simply contains two MultiLayerPerceptrons as encoder and decoder.