Samples: Add DLPack GPU interop sample and stitching ROI masking

[csu-bot-zivid]

New sample: DLPack GPU interop

source/applications/advanced/cuda/capture_and_convert_to_dlpack_tensor_on_cuda.py
shows how to hand a Zivid DeviceArray to PyTorch on the GPU without a
CPU round-trip. Two paths are demonstrated:

1. Recommended: torch.as_tensor(device_array, device="cuda"), which
   imports the buffer through the CUDA array interface that DeviceArray
   already implements — no pointer, shape or stride handling needed.
2. Optional: build a DLPack capsule directly from the DeviceArray's
   device pointer, shape, strides (in elements), element data type and
   CUDA device, exposed through the standard __dlpack__ /
   __dlpack_device__ protocol and imported with torch.from_dlpack.
   Use this when your consumer expects a DLPack capsule directly. The
   DeviceArray is kept alive until the consumer releases the tensor.

Both hand-offs are verified zero-copy by comparing data pointers, and
both are scoped to the CUDA backend. The sample requires a CUDA-capable
GPU and PyTorch with CUDA support (pip install torch); PyTorch is a
third-party dependency that Zivid does not provide or support.

The README now lists the sample among the cuda samples.

Stitching GUI: region of interest masking

The stitching verification tab now has an optional Region of Interest
box for keeping background clutter out of the stitched point cloud.

• Tick Enable ROI masking and set min/max X, Y and Z extents in mm.
• The extents are given in the coordinate frame that stays fixed
  relative to the scanned object: the robot base frame for eye-in-hand
  and the robot flange frame for eye-to-hand. The group box title states
  which frame is currently in use.
• Masking is applied to each capture before the hand-eye transform, both
  for captures taken in the GUI and for captures loaded from disk.
• If the box would remove every point of a capture, that capture is
  shown unmasked and a warning asks you to adjust the extents, so you
  never end up with a silently empty view.