[Afloat16]

What does this PR do?

Fixes #6544.

OneCycleLR captures optional arguments in kwargs but did not forward them to PyTorch. The documented three_phase=True option therefore silently used the two-phase curve, changing both learning rates and cycled momentum. total_steps=None with epochs and steps_per_epoch also failed to construct the documented inferred-length schedule.

Forward the optional arguments to the native scheduler. Explicit wrapper arguments and the maximum learning rates taken from optimizer groups remain unchanged. Unknown optional arguments now receive the native constructor's validation.

Validation

• 15 native regression cases: upstream base 11 failed / 4 passed; patch 15 passed.
• Two real SGD/Adam parameter groups are compared step by step against torch.optim.lr_scheduler.OneCycleLR for learning rate, momentum/betas and parameter updates, using cosine/linear annealing and two/three phases.
• Epoch-derived total steps, real optimizer/scheduler checkpoint continuation and an unknown-argument control are covered.
• python -m pytest tests/test_optimizer/test_onecycle_options.py tests/test_optimizer/test_lr_scheduler.py -q: 16 passed, including the existing scheduler save/load regression.
• Applicable configured pre-commit hooks and git diff --check passed.
• Python 3.12 and PyTorch 2.9.1+cpu. GPU/distributed training and the full repository suite were not run.