utils/cv2: use rsplit in cv2_load_image so Windows drive-letter paths work

[narr2atnarrable]

Summary

cv2_load_image in tracklab/utils/cv2.py parses a vid://<path>:<frame_id> URI with str.split(":"). On Windows, paths contain a drive-letter colon (e.g. C:/Users/.../clip.mp4), so the URI vid://C:/Users/me/clip.mp4:0 splits into three pieces and the tuple unpack raises ValueError.

Using rsplit(":", 1) cleaves only the trailing :<frame_id> and leaves the rest of the path intact. It works identically on POSIX (no change in behaviour) and fixes Windows.

Reproduction (POSIX shell, same bug a Windows user hits)

>>> "C:/Users/me/clip.mp4:0".split(":")
['C', '/Users/me/clip.mp4', '0']          # 3 pieces → ValueError on unpack
>>> "C:/Users/me/clip.mp4:0".rsplit(":", 1)
['C:/Users/me/clip.mp4', '0']             # correct

Any Windows user loading frames through the vid:// scheme (ExternalVideo wrapper, dataset=video configs, the video reader path in image-level pipelines) hits it:

File ".../tracklab/utils/cv2.py", line 59, in cv2_load_image
    video_file, frame_id = file_path.split(":")
ValueError: too many values to unpack (expected 2)

Change

One line in tracklab/utils/cv2.py:

-        video_file, frame_id = file_path.split(":")
+        video_file, frame_id = file_path.rsplit(":", 1)

Risk

None on POSIX — rsplit(":", 1) on a single-colon string returns the same 2-tuple as split(":"). The only behavioural change is that multi-colon strings now parse as "everything up to the last colon is the path, the tail is the frame id", which is the intended semantics.

Context

Encountered while running sn-gamestate on a workstation pair (Linux + Windows) for a Blackwell-GPU compatibility sweep. Filing alongside a sibling PR to SoccerNet/sn-gamestate for a related pandas-2 .iloc[0] fix; both were needed to get the pipeline end-to-end on a mixed-OS setup.