Fix PytorchLARS updates when momentum is zero

[mottopanikeiku]

Fixes #2079.

I use the weight-decayed gradient when PytorchLARS has zero momentum, keeping the existing update-norm limit. This fixes the default step() crash and prevents a zero-momentum group from reusing the previous parameter's momentum update. The native LARS variants are unchanged.

Validation on CPU with Python 3.12 and PyTorch 2.8.0+cpu:

• OMP_NUM_THREADS=2 python -m pytest tests/test_optim.py -k lars -q: 10 passed, 4 existing bfloat16 skips.
• Both new regressions fail before the fix.
• A 16-step training smoke matches PyTorch SGD with mixed momentum groups, weight decay, Nesterov momentum, and closures.
• pre-commit run --all-files: all hooks passed.

I used OpenAI Codex for investigation, implementation, and testing. This has been manually reviewed.