feat: support fsdp2 muon optimizer

[RangiLyu]

[nil0x9]

It might be more desirable to introduce some mechanism to separate fused params IMHO? For example, MoE projections are implemented with GroupedLinear in XTuner, where 2D param actually means a bundle of linear weights. In this case, it might be more beneficial to split this single param tensor and treat them as different projections in NS iteration.