Implement GPU acceleration support across optical flow, face detection, and pose estimation

[Copilot]

GPU (CUDA) acceleration was only available for pose estimation. This adds opt-in GPU support to dense/sparse optical flow and face blurring, with automatic CPU fallback when CUDA is unavailable.

Changes

Shared utility

• _utils.py: Added get_cuda_device_count() — wraps cv2.cuda.getCudaEnabledDeviceCount() safely, returns 0 on failure. Exported from package root.

GPU-accelerated functions

• flow.dense(use_gpu=True): Uses cv2.cuda.FarnebackOpticalFlow when available. Pre-allocates two cv2.cuda_GpuMat objects and swaps references each frame to avoid per-frame GPU memory allocations.
• flow.sparse(use_gpu=True): Uses cv2.cuda.SparsePyrLKOpticalFlow with the same GpuMat reuse strategy for both frame buffers and tracked point arrays.
• blur_faces(use_gpu=True) / CenterFace(use_gpu=False): Sets DNN_BACKEND_CUDA / DNN_TARGET_CUDA on the ONNX network when CUDA is detected.
• pose(device='gpu'): Refactored to use the shared get_cuda_device_count() instead of inline logic.

All GPU paths fall back to CPU with an informative message if cv2.cuda.FarnebackOpticalFlow / SparsePyrLKOpticalFlow are absent (require opencv-contrib built with CUDA) or no CUDA devices are found.

Usage

import musicalgestures as mg

# Check GPU availability
print(mg.get_cuda_device_count())  # 0 on CPU-only machines

mg_video = mg.MgVideo("video.mp4")
mg_video.flow.dense(use_gpu=True)   # GPU Farneback, falls back to CPU
mg_video.flow.sparse(use_gpu=True)  # GPU LK, falls back to CPU
mg_video.blur_faces(use_gpu=True)   # GPU ONNX inference, falls back to CPU