[Feature] Add --mg-fl-prefer argument for Megatron-LM-FL vendor selection

[lihongyang1990]

Summary

• Add --mg-fl-prefer CLI argument to select the preferred vendor backend (cuda/musa/txda) for Megatron-LM-FL plugin override dispatch
• Sync the argument value to the MG_FL_PREFER environment variable at training startup, following the same pattern as --te-fl-prefer / TE_FL_PREFER
• Rename the argument group from "flagscale transformer engine fl" to "flagscale fl" to reflect the broader scope

Changes

flagscale/train/megatron/training/arguments_fs.py

• Add --mg-fl-prefer argument with choices ['cuda', 'musa', 'txda'], default empty string
• Rename argument group title to "flagscale fl"

flagscale/train/megatron/training/training.py

• Sync args.mg_fl_prefer to os.environ['MG_FL_PREFER'] in pretrain(), alongside the existing TE_FL_PREFER sync logic

Usage

CLI:

python train.py --mg-fl-prefer musa --te-fl-prefer reference

YAML config:

model:
  mg_fl_prefer: musa
  te_fl_prefer: reference

Test plan

• Verify --mg-fl-prefer musa sets MG_FL_PREFER=musa in the environment
• Verify omitting --mg-fl-prefer does not set MG_FL_PREFER
• Verify invalid values are rejected by argparse choices validation
• Verify compatibility with existing --te-fl-prefer argument

[lxd-cumt]

LGTM