- update to 4.9.0,  highlights below, for details check
  https://github.com/opencv/opencv/wiki/ChangeLog#version490
  Highlights of this release:
  * Core Module:
    + Added cv::broadcast
    + Fixed several rounding issues on ARM platform
    + Added detection & dispatching of some modern NEON
      instructions (NEON_FP16, NEON_BF16)
    + Added optimization for LoongArch 128-bit vector,
      detection & dispatching of LoongArch
  * DNN module patches:
    + Experimental transformers support
    + ONNX Attention layer support
    + ONNX Einsum layer support
    + OpenVINO backend for INT8 models
    + ONNX Gather Elements layer
    + ONNX InstanceNorm layer
    + Better support of ONNX Expand layer with cv::broadcast
    + Improved DNN graph fusion with shared nodes and commutative
      operations
    + New fastGEMM implementation and several layers on top of it
    + Winograd fp16 optimizations on ARM
    + Tests and multiple fixes for Yolo family models support
    + New layers support and bug fixes in CUDA backend: GEMM,
      Gelu, Add
    + CANN backend: bug fix, support HardSwish, 
      LayerNormalization and InstanceNormalization
    + LayerNormalization: support OpenVINO, OpenCL and 
      CUDA backend.
  * G-API module:
    + Intel® OpenVINO™ DL inference backend:
      - Introduced "inferenence only" ("benchmark") mode 
        in the OV2.0 backend.
      - Fixed model layout setting issue in the OV2.0 backend.
      - Fixed/relaxed various asserts in the OV2.0 backend.
    + Core and image processing functionality:
      - Fluid kernels were rewritten to new universal intrinsics.
        Thanks for this contribution!
    + Streaming and video functionality:
      - Introduced a QueueSource: an alternative way to manually
        push input frames to the G-API pipeline in the streaming
        mode.
      - Introduced VAS Object Tracker (OT) for the various video
        analytics scenarios.
    +  Python bindings:
      - Exposed VAS OT in G-API Python bindings.
    + Other changes and fixes:
      - Updated ADE (the G-API's graph library) to the latest
        version.
      - Various code clean-ups and warning fixes.
  * Objdetect module:
    + Implemented own QR code decoder as replacement for QUIRC
      library
    + Bug fixes in QR code encoder version estimation
    + More accurate Aruco marker corner refinement with
      dynamic window
    + Fixed contour filtering in ArUco
    + QR code detection sample for Android
    + Multiple local bug fixes and documentation update for
      Aruco makers, Charuco boards and QR codes.
  * Video:
    + Google Summer of Code: added a new object tracking API
      TrackerVit for a vision transformer-based VitTrack. 
      This work is done by LIU Pengyu.
  * VideoIO:
    + Videoio: Add raw encoded video stream encapsulation
      to cv::VideoWriter with CAP_FFMPEG
    + Fix GStreamer backend with manual pipelines.
  * Calibration module:
    + Multiple fixes and improvements chess board calibration
      rig detector.
    + calibrateCamera throws exception, if calibration system
      is underconstrained.
    + Fixed bug in findEssentialMat with USAC
    + Fixed out-of-image access in cv::cornerSubPix
    + Fixed crash in ap3p
    + Fixed stereoRectify image boundaries
    + Fixed "use after free" issue in essential_solver.cpp
  * Python Bindings:
    + Added type stub generation for missed types and manually
      wrapped types.
    + Added read-only flag handling for Numpy arrays.
    + Fixed exception handling and bindings for in module.
    + Improved error messages in Numpy array type handling.
    + Fixed constructors documentation in Python.
  * Platforms and hardware Support:
    + Experimental CUDA support as first class language in CMake
    + Added experimental support for Apple VisionOS platform
    + Add support Orbbec Gemini2 and Gemini2 XL camera
    + Fix fullscreen behavior on macOS
  * Other:
    + OpenCV Summer of Code: semi-automated refactoring across
      multiple pull requests by HAN Liutong made our CPU-optimized
      code compatible with SIMD with variable vector
      length (RISC-V RVV)

• Files Changed
• Browse Source