fix: guard temp-dir cleanup so image/dir input doesn't crash main()

[MushiSenpai]

Summary

scripts/inference.py crashes during cleanup when the input is an image or an
image directory (not a video).

save_dir_full is assigned only on the video-input branch:

if get_file_type(video_path) == "video":
    save_dir_full = os.path.join(temp_dir, input_basename)
    ...
elif get_file_type(video_path) == "image":   # save_dir_full never set
    ...
elif os.path.isdir(video_path):              # save_dir_full never set
    ...

…but cleanup calls it unconditionally:

shutil.rmtree(save_dir_full)   # UnboundLocalError for image / dir input

So for image or directory input, main() raises
UnboundLocalError: cannot access local variable 'save_dir_full' at cleanup. It's
swallowed by the surrounding except Exception, but only after the output video
is already written — so the user sees a spurious "Error occurred during processing:
…" and the temp frame directory / coord file are left uncleaned.

The fix

Initialize save_dir_full = None before the input-type branch and guard the cleanup:

save_dir_full = None  # only set for video input
...
if save_dir_full is not None:
    shutil.rmtree(save_dir_full)

Two lines, no behavior change for video input.

Validation

The bug is pure control flow. Reproducing inference.py's exact structure
(save_dir_full bound only in the video branch, referenced at cleanup):

Input	Before	After
video	cleanup runs	cleanup runs (unchanged)
image	UnboundLocalError	cleanup skipped cleanly

Scope

Self-contained correctness fix, independent of #419 (the PyTorch 2.6 weights_only
face-parse fix). No existing issue tracked this; happy to open one if maintainers prefer.

---

Disclosure: this fix was drafted with AI assistance (Claude) and reviewed, tested, and
submitted by me. The commit carries a Co-Authored-By trailer.