Fix Megatron checkpoint save and resume state

[nightlessbaron]

What changed

• Pass Megatron's preprocess_common_state_dict hook when saving non-LoRA checkpoints.
• Avoid advancing the optimizer scheduler a second time after a normal checkpoint resume.
• Continue reconstructing scheduler position for finetuning and --no-load-optim, where scheduler state is intentionally not restored.
• Add four focused regression cases.

Why

A normal resume already restores scheduler state in load_checkpoint; stepping it again double-counts completed iterations. Checkpoint saves also need Megatron's common-state preprocessing to produce the expected distributed checkpoint state.

Validation

• Focused checkpoint tests: 4 passed.
• Black: passed.
• Ruff: passed.
• py_compile: passed.
• git diff --check: passed.

Related Megatron optimizer-state fix: LLM360/Megatron-LM#6.