Decay OneCycle momentum instead of growing it past 1.0

[vineethsaivs]

With cycle_momentum on and any decay_mom_rate, OneCycle grows the first Adam beta without bound in its decay phase and pushes it past 1.0.

_get_decay_mom multiplies cycle_max_mom by (1 + rate * interval) while _get_decay_lr divides the learning rate by the same factor. Both the docstring and the command line help call the parameter a decay rate. Divide in _get_decay_mom so momentum decays from the cycle maximum, mirroring _get_decay_lr.

Test: TORCHDYNAMO_DISABLE=1 DS_ACCELERATOR=cpu PYTHONPATH=. python -m pytest tests/unit/runtime/test_lr_schedulers.py -k 'one_cycle or momentum' -q. A regression fails before, betas[0] reaching 1.045 at step 30 and rising, a value torch.optim.Adam rejects in its own constructor. Fifteen CPU tests pass after against hand derived values. An existing assertion that pinned the growth now expects the decay. Changed-file pre-commit passes. Executed on Apple M2 Pro CPU; no CUDA or distributed training run. The DistributedTest failures in this file are pre-existing on unpatched source here.

[vineethsaivs]

Could you approve the CI run? A regression fails before, betas[0] reaching 1.045 at step 30 and rising, a value torch.optim.Adam rejects in its own constructor. Fifteen CPU tests pass after against hand derived values. An existing assertion that pinned the growth now expects the decay. Changed-file pre-commit passes.