Fix eval mode for wrapped optimizers

[uuzzrm]

What does this PR do?

Fixes #4159.

AcceleratedOptimizer.train() already reaches through the extra optimizer wrapper used by DeepSpeed. eval() only checked the outer wrapper, so schedule-free optimizers could remain in training mode when an evaluation checkpoint was saved.

This adds the matching one-level fallback and a regression test using a small schedule-free-like optimizer wrapped in a DeepSpeed-like object.

Testing

• PYTHONPATH=src uv run --with torch --with pytest --with numpy --with packaging --with psutil --with pyyaml --with huggingface_hub --with safetensors --with '.[test_prod]' pytest tests/test_optimizer.py -q
• uv run --with ruff ruff format --check tests/test_optimizer.py src/accelerate/optimizer.py
• git diff --check

AI writing disclosure

• AI-assisted: Codex helped investigate the issue, draft the focused patch and regression test, and run the checks. I reviewed the changed lines, reproduced the failure before the fix, and verified the passing test result.
• AI-generated
• No AI usage

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• I read the contributor guideline and searched existing issues and PRs before starting.
• I linked the issue addressed by this change.
• Documentation needs updating.
• I added a regression test for the changed behavior.