[Afloat16]

What does this PR do?

Fixes #6540.

LinearWarmupLR currently continues linearly below zero after total_steps. An extra training step or continuation of a completed schedule therefore reverses parameter updates. For total_steps=4 and warmup_steps=1, the scheduler reaches zero at step 4 and then returns negative rates.

Clamp the decay numerator at zero, preserving the existing warmup and decay trajectory through the configured horizon. Completed schedules now keep a zero learning rate with both ordinary and explicit epoch stepping.

Validation

• 17 native regression cases: upstream base 14 failed / 3 passed; patch 17 passed.
• Hard-coded learning-rate trajectories cover zero/one/three warmup steps and both stepping forms; two real SGD parameter groups with momentum are compared with independent manually assigned reference rates.
• Explicit epoch jumps and real optimizer/scheduler checkpoint round trips before and after completion verify the zero tail.
• python -m pytest tests/test_optimizer/test_linear_warmup_floor.py tests/test_optimizer/test_lr_scheduler.py -q: 18 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. Validation covers valid configurations with warmup_steps < total_steps.