ModelAssembler, torchModel and saving/loading of parameters

[PeterCalifano]

SaveTorchModel function has troubles in saving traced models when the latter is composed by multiple piece stacked through nn.Sequential. Furthermore, this method is not able to move all modules to the target device and requires .to() method to be overridden and explicitly called on some parts. This is also indicated by torchsummary, which does not see details of one of the modules. Activities to perform:

• Investigate issue of tracing models composed of multiple modules.
• Investigate issue of .to() method trying alternative stacking other than nn.Sequential.
• From task 2, define ModelAssembler as class to produce assembly of models (does not need to be too flexible for now) or to be used directly as nn.Module itself.
• Define ways to load or extract subsets of model from ModelAssembler to improve flexibility and saving of experiments.
• From task 1 and 2, define torchModel wrapper providing overriding of necessary methods (e.g. to()) to handle tracing, reloading and full compatibility with torch.