Fix checkpoint path bug, mask channel inconsistency, and a few related cleanups

[kyroX-D]

Hey! Was reading through the codebase and noticed a few things that
looked off figured I'd fix them and send a PR rather than just sit
on it. Happy to split this up if you'd prefer smaller PRs, just let me know.

bugs

CHECKPOINT_DIR was defined twice in CorridorKeyModule/backend.py
The absolute path at the top of the file was silently shadowed by a
relative redefinition further down. This broke model lookup and the
auto-download path whenever the tool was launched from a directory
other than the repo root. Removed the second definition.

Video alpha masks were read from the wrong channel
clip_manager.py and backend/frame_io.py pulled BGR[2] (red) out of
video mask frames, while image masks went through the blue channel /
grayscale path elsewhere. For grayscale-encoded sources the two paths
happen to agree, but for any color-encoded mask they silently disagree.
Both video paths now go through cv2.cvtColor(BGR2GRAY), which also
matches what the existing VideoMaMa mask pipeline does. Also fixed the
docstring in frame_io.py that claimed index 2 in BGR was the blue
channel it's red.

Frame sequences were sorted lexicographically in the CLI path
For non-zero-padded sequences (frame_2.png, frame_10.png) that puts
frame_10 before frame_2, so frames got processed out of order.
backend/clip_state.py already used natural sort via natsorted();
the CLI just hadn't caught up. Both paths now use the same helper from
backend/natural_sort.py.

Cleanup branch in the failed MLX download pointed at the wrong path
When the download failed mid-stream, the code tried to remove model_path
but the partial file lived under cache_path (.tmp). The wrong file
was checked, so the leftover .tmp was never actually cleaned up.

Cleanups

• Dropped the leftover __main__ block at the bottom of clip_manager.py.
  It raised NotImplementedError for the wizard action and was fully
  superseded by corridorkey_cli.py.
• Removed a duplicate import numpy as np from inside run_inference()
  already imported at module level.
• Replaced three traceback.print_exc() sites with logger.exception()
  so stack traces flow through RichHandler and respect the configured
  logging setup instead of bypassing it onto stderr.
• Moved the ffmpeg_tools import out of the per-frame loop up to the
  module header where it belongs.
• Added pytest-env to the dev dependency group. pyproject.toml
  already declared env = [...] under [tool.pytest.ini_options] to
  set OPENCV_IO_ENABLE_OPENEXR=1, but the option was silently ignored
  without pytest-env installed, so EXR-related tests failed locally
  on Windows.
• A few minor lint/style tweaks via ruff --fix (unused argparse
  import, sorted import block, trailing whitespace).

Testing

uv run pytest tests/ — 363 passed, 5 skipped (GPU/MLX-only, expected).
On Windows the EXR tests pass too once pytest-env is installed via
uv sync.

Let me know if anything looks weird or you want anything reworked!