Weight dropout

[albertz]

Dropout on the weights.

This needs to depend (conditional logic, #24) on the train flag (#18). Both is implemented now, and we can just use:

with nn.Cond(nn.train_flag()):
  ...

Edit Actually not really because this is already internal to nn.dropout.

Also see nn.stochastic_depth as another example using nn.Cond with nn.train_flag.

Very related is also variational parameter noise (#240).

[albertz]

This is implemented now, see nn.weight_dropout.