Support model "buffers"

[Kaixhin]

Basically these are parameters that aren't updated via gradient descent (but would be serialized - a good example that already exists here is the running mean or running variance in batch normalisation); they could also be completely fixed.

The current working solution is to just add Variables that don't need gradients within initialize_parameters. We need to also make sure optimizers skips anything with requires_grad=false.

[jgehring]

Note that you can serialize custom data in your model by implementing a custom serialize or load/save function.