Fix LR schedulers when warmup/decay rounds to zero

[xsinxx]

Summary This PR makes LR schedulers robust to warmup=0 / decay=0 (and fraction-based warmup/decay that rounds to 0). Previously, these configurations could lead to division-by-zero in the linear warmup/decay helpers.

Changes

• Treat warmup=0 as “no warmup” and avoid entering linear warmup when warmup == 0 :

  • ConstantWithWarmup
  • WSD

• Treat decay=0 as “no decay” and avoid entering linear decay when decay == 0 :

  • WSD
  • CosWithWarmupAndLinearDecay

• Handle WSDS period decay steps D == 0 by returning decay_min_lr directly to avoid division-by-zero.
  Tests

• Added unit tests covering:

  • ConstantWithWarmup(warmup=0)
  • ConstantWithWarmup(warmup_fraction=...) that rounds to zero warmup
  • WSD(warmup=0, decay=0)

• Ran: python -m pytest -q src/test/optim/scheduler_test.py (all passed)
  Behavioral Impact

• No behavior change for configurations with warmup > 0 and decay > 0 .

• Configurations with warmup=0 / decay=0 now behave as “disabled phase” instead of raising due to division-by-zero.