[config] Reject unknown optimizer names in ParamGroupConfig

[YeonwooSung]

Summary

ParamGroupConfig.optimizer_name accepted any string. A typo such as Adamw became NotImplementedError only after the model was built.

The allowed set is still Adam, AdamW, and DistMuon, now as a single module-level factory map. Unknown names raise ValueError when the param group is constructed.

Test plan

• New tests construct ParamGroupConfig only (no model / container): accept the three names; reject Adamw and foo
• CI CPU unit tests for test_optimizer_param_groups.py

[tianyu-l]

instead, we should

• create torchtitan native Optimizer(Configurable, torch.optim.Optimizer)
• create torchtitan native AdamW and DistMuon (https://github.com/pytorch/torchtitan/blob/main/torchtitan/distributed/flex_shard/dist_muon.py#L252) -- the AdamW one can be done via a from_torch helper.
• here specify optimizer: Optimizer.Config

if it sounds complicated, leave it to me.