Training package export

The package declares its scope in a docstring and performs one explicit re-export source L1–L2:

from .optimizers import MuonAdamW as MuonAdamW

The alias is the same class object imported from training/optimizers.py, not a second optimizer implementation. It exposes neither TrainConfig nor the training loop, which are imported directly from training/trainer.py. There is no training configuration, I/O or mutation of model parameters in this initializer.