fix: allow local LoRA paths in offline mode

[pctablet505]

Fixes #14195

Moves the local file and directory checks in _best_guess_weight_name() before the offline-mode guard, so explicit local LoRA files and directories can be loaded when local_files_only=True or HF_HUB_OFFLINE=1 is set. Remote repository IDs without an explicit weight_name continue to raise in offline mode, and model_info() is still not called.

[fropych]

Thanks for looking into this. I opened #14195 to coordinate the fix and already have a local branch and commit ready, including focused regression tests. This would be my first open-source contribution, so I was hoping to take the implementation through the PR and review process myself.
Since the issue has not yet received maintainer acknowledgment, @sayakpaul, could you please advise whether I should open my PR or coordinate with #14198? I’m happy to collaborate and would like to avoid duplicate work.

[sayakpaul]

@pctablet505 can you maintain the PR template and acknowledge how much AI assistance you had in this PR?

[fropych]

@sayakpaul, I would like to clarify the coordination history here.
In #14195, I explicitly said that I wanted to implement this fix and asked for confirmation before opening a PR. Since my work was AI-assisted, I followed the contribution guide and waited for maintainer acknowledgment.
While I was waiting, #14198 was opened by another contributor with essentially the same fix, without coordinating with me.
I have now opened #14204. This would be my first open-source contribution, and I was hoping to take it through the review process myself.
My PR:

• follows the full PR template
• clearly discloses the AI assistance
• includes the self-review notes and actual check results
• covers every regression case described in the issue
• verifies that model_info() is not called in offline mode
• preserves the existing behavior for directories with no matching files or multiple matches.

One public-API test in #14198 changes os.environ after HF_HUB_OFFLINE has already been imported. As a result, it does not actually exercise that offline path. My test patches the imported constant directly.
I’m happy to collaborate, but I would appreciate it if the coordination history and test coverage could be considered when deciding which PR should move forward.