fix: use launch world size for CLI scheduler steps

[dorodo0721]

Summary

• Resolve the effective world size using the existing external-launcher detection: local launches use CLI nprocs, while external launches require a positive integer WORLD_SIZE.
• Use the resolved world size only when calculating CLI scheduler total steps, so warmup and decay lengths reflect global data-parallel sharding.
• Preserve TrainConfig.nprocs, launcher behavior, and existing configuration validation; update the distributed training guide.

Tests

• Add launch-world-size resolution tests and CLI-to-scheduler regression coverage for cosine, SGDR, and WSD, including mismatched CLI and environment values.
• Before CLI wiring: regression tests reproduced the bug (9 failed, 6 passed). After the fix: targeted tests passed (38 passed, 2 deselected).
• Full CPU test suite: python -m pytest -q tests/ --tb=short resulted in 712 passed, 50 skipped, and 2 dependency deprecation warnings. Tests requiring localhost sockets were rerun successfully outside the sandbox.
• Ruff import checks, formatting checks, and git diff --check passed.

Limitations

• Validated on macOS ARM64 with Python 3.12.13 and PyTorch 2.11.0.
• No actual GPU training or multi-node torchrun execution; CLI regression tests simulate launcher environment variables and instantiate real schedulers without starting training.