Fix invalid group reference in revert_checkpoint_conversion_mapping

[A-S-Manoj]

Description

revert_checkpoint_conversion_mapping unconditionally stripped the capture
group out of each source regex pattern, even when the corresponding target
pattern still referenced that group via a backreference (e.g. \1).

This is fine for mappings where the target is a plain string, but
transformers-5.x checkpoint_conversion_mappings (e.g. Qwen3.5/3.6-MoE)
include entries like:

{r"^model\.language_model\.(.+)$": r"model.\1"}

Stripping the group here leaves \1 in the target dangling, and
re.subn(source_pattern, target_pattern, name) raises:

re.error: invalid group reference 1 at position 7

The fix only strips the capture group when the target pattern does not
backreference it, so patterns that rely on the captured suffix keep working
while the existing bare-tensor-suffix handling is unaffected.

Type of Change

Bug fix

Related Issues

Fixes #2028

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.
• The CUDA CI has passed. You can trigger it by commenting /azp run Unit-Test-CUDA-AutoRound.

[azure-pipelines]

Commenter does not have sufficient privileges for PR 2035 in repo intel/auto-round

[yiliu30]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).