fix: make post-warmup cosine LR scheduling effective and configurable

[yamsam]

Summary

Fix the learning-rate schedule so that cosine decay is applied correctly after warmup.

The previous CosineAnnealingWarmUpRestarts implementation used
MultiplicativeLR(lambda=1.0) after warmup. Despite its name, it therefore kept
the learning rate constant instead of applying cosine annealing.

This PR introduces a real linear-warmup → cosine-decay schedule while preserving
the previous behavior for backward-compatible experiments.

Based on dev.

Changes

Scheduler behavior

• Add WarmupCosineAnnealingLR: linear warmup followed by
  CosineAnnealingLR.
• Rename the previous warmup → constant implementation to
  WarmupConstantLR.
• Keep CosineAnnealingWarmUpRestarts as an alias for the legacy constant
  implementation to preserve existing imports.
• Add build_lr_scheduler(...) to select cosine or constant.
• Prevent the legacy final-10-epoch manual LR step-down from overriding the
  cosine schedule. The legacy step-down remains active only in constant mode.

Resume behavior

• Rebuild and fast-forward the scheduler to the completed training progress
  when resuming from a checkpoint.
• Avoid resetting a resumed cosine run directly to the peak learning rate.
• Preserve schedule continuity when resuming with either the same or a new base
  learning rate.
• Convert completed epochs to optimizer steps when using batch-level scheduler
  updates.

Configuration

The following options are available through TrainConfig and
train_predictor.py:

Option	Default	Description
lr_schedule_type	cosine	cosine or legacy constant
learning_rate	1e-4	Peak/base learning rate
warm_up_epoch	5	Warmup duration in epochs
lr_start_factor	0.1	Initial LR as a fraction of the peak LR
lr_eta_min	1e-6	Minimum cosine LR
lr_cosine_t_max	0	Cosine duration in epochs; 0 selects the remaining epochs automatically
lr_scheduler_interval	epoch	Update the scheduler per epoch or per batch

When lr_scheduler_interval=batch, warmup duration, explicit T_max, and
resume progress are converted from epochs to optimizer steps using the training
DataLoader length.

The legacy final-phase parameters remain unchanged.

Behavior change and compatibility

• New training runs use actual cosine decay by default.
• The previous training policy remains available with
  --lr_schedule_type constant.
• Existing imports of CosineAnnealingWarmUpRestarts continue to work and
  retain the previous warmup → constant scheduler behavior.
• To reproduce an older run when resuming, explicitly select
  --lr_schedule_type constant.

Verification

• LR scheduler tests: 15 passed
• Ruff lint: passed
• Ruff format check: passed
• Pre-commit hooks: passed
• Verified cosine and legacy constant behavior with the mini dataset.
• Verified checkpoint resume without an LR jump.
• Verified custom start_factor, eta_min, T_max, and batch-level scheduler
  updates with a mini-dataset training run.

[yamsam]

Comparison results with the legacy implementation