chore(synthesis): bundle MediaPipe deps + hand_landmarker model for HandGestureDetector block

[harsh839]

Summary

Completes the backend side of the MediaPipe HandGestureDetector custom block
for the Build & Download flow, addressing the remaining bundling gap in
#470 (the block itself, HandGestureDetector.vc3 + hand_landmarker.task, is
already merged upstream via VisualCircuit-resources#17).

When a circuit that uses the hand-gesture block is built, the generated ZIP must
ship its third-party runtime and model. This PR wires that up by mirroring the
exact pattern already used for FaceDetector (haar) and ObjectDetector
(yolov3):

1. requirements.txt — added
   'HandGestureDetector': ['mediapipe', 'opencv-python', 'numpy'] to
   BLOCK_DEPENDENCIES. mediapipe has no entry in COMMON_PIP_ALIASES, so the
   AST sweep alone would only add an unresolved mediapipe name; listing it here
   guarantees the real pip package lands in the generated requirements file.
2. Model bundling — added
   'HandGestureDetector': 'utils/models/hand_landmarker/**/*' to
   OPTIONAL_FILES, mirroring how FaceDetector/ObjectDetector pull their
   model assets into the ZIP (via the paths_to_include whitelist in
   syntheize_executioner, keyed off the optional_files[script_name] flag set
   in syntheize_modules).
3. Model asset — ships
   backend/staticfiles/synthesis/utils/models/hand_landmarker/hand_landmarker.task
   (7.6 MB, from Google's official MediaPipe model zoo — same convention as the
   committed yolov3-tiny.weights), plus a README.md noting source/license.

Verification

Backend CI does not run tests, so correctness is verified locally:

• python -m py_compile backend/synthesis/synthesis.py — passes.
• Reimplemented the bundler's exact include/exclude matching
  (Path.relative_to('synthesis').match(...) against the OPTIONAL_FILES
  globs, as done in syntheize_executioner) against the on-disk
  hand_landmarker.task path:
  • glob utils/models/hand_landmarker/**/* includes the model when the
    block is present (True),
  • it is not swallowed by the FaceDetector/ObjectDetector exclude globs
    (True),
  • resolved deps = ['mediapipe', 'opencv-python', 'numpy'].
• git diff origin/master...HEAD touches only synthesis.py (2 dict lines) and
  the new model directory — no unrelated churn.

Notes

• Per the maintainer's guidance in New CV block proposal: MediaPipe Hand & Gesture detector #470, this covers backend bundling; the
  frontend block registration already shipped in VisualCircuit-resources#17.
• Independent of fix(outputs): size shared-memory wires from promoted dtype #507 (outputs wire-share fix) and chore(backend): format Python codebase with black #508 (black formatting),
  since this branch was cut from current master (7a42b74).

Closes the bundling portion of #470.

[harsh839]

@jmplaza could you take a look when you get a chance? This bundles the MediaPipe hand_landmarker model so the HandGestureDetector block's synthesized project ZIP includes the model file (see OPTIONAL_FILES / BLOCK_DEPENDENCIES in backend/synthesis/synthesis.py).