[FEATURE] Add Support for Muon Optimizer

[caojiaolong]

The recently introduced [Muon Optimizer](https://github.com/KellerJordan/Muon) has shown promising results, claiming to outperform commonly used optimizers such as AdamW, Shampoo, and SOAP ([reference benchmarks](https://github.com/KellerJordan/modded-nanogpt/tree/master/records/102924_Optimizers)).

Given its potential advantages, it would be valuable to have Muon integrated into the timm library for broader experimentation and adoption.

Is there any plan or interest in adding support for Muon Optimizer to timm?

[rwightman]

@caojiaolong I was pondering doing it, but it's in torch now https://github.com/pytorch/pytorch/blob/main/torch/optim/_muon.py ... so that lowers priority for adding to timm unless there are any modes/features not in the torch version. EDIT: I will add the mapping to timm to allow using the torch.optim.Muon w/ the timm factory.

One thing with the original, I had some q re compatibility with convnet weight layouts as the dim handling isn't always friendly with convs and transpose convs. Think some of that might have been fixed, I have to try out the torch version, haven't had a chance.

[rwightman]

I could never get Shampoo and SOAP working well (finding decent hparams for vision tasks). Though another related optimizer Kron works quite well.

[rwightman]

@caojiaolong looking more closely, surprisingly the pytorch impl of Muon only supports 2D parameters... sooo guess that's not going to work. I figured they'd generalize it better if it was going to be included there...