LARS initialization is different than pytorch Optimizer in documentation

[zlenyk]

Thanks for creating this repo! It helps me a lot during training.
I wanted to ask/suggest if it was possible the change the initialization of the optimizer to be consistent with pytorch.optim.Optimizer.
In documentation (https://pytorch.org/docs/stable/optim.html) it says that optimizer should be given iterable with params during initialization. Because it's not like that right now, it makes it impossible to use this implementation with for example horovod DistributedOptimizer.

Thank you!