Fix percentage formatting in help messages for gradient checkpointing, Liger Kernel, and empty cache steps

[qgallouedec]

In huggingface/trl#5112 a user reported that trl sft --help fails

It's because three inherited args from TrainingArguments (torch_empty_cache_steps, gradient_checkpointing and use_liger_kernel)help strings include bare % which break argparse formatting on Python 3.13. This PR fixes it.

[stevhliu]

thanks for fixing! this should also close #43980

[qgallouedec]

ah right sorry for the duplicate

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[albertvillanova]

Thanks, glad to see this has been already fixed!

For future reference, in most OSS workflows the usual practice is to keep the earlier PR open and mark the later one as a duplicate before closing it. That helps preserve the original discussion context and avoids losing review history on the first contribution.

Not a big deal at all, just mentioning it for consistency going forward.