Support weight reparametrization

[kitterion]

Feature description

It would be nice if burn supported some way to apply weight reparametrization to layers like found in pytorch (e.g. weight_norm, spectral_norm).

Feature motivation

pytorch docs provide a decent explanation and relevant links but in short, this can stabilize training of certain networks.

(Optional) Suggest a Solution

In the current design this kind of api probably belongs on Param since dynamically replacing fields, like pytorch does, is not really feasible. Param would have to switch out implementations based on a particular reparametrization.

[nathanielsimard]

You can use a visitor pattern to apply a transformation to each weight, similar to how the optimizer works.