Samples: Add DLPack CUDA sample and CSV output for ZividBenchmark

[csu-bot-zivid]

New sample: DLPack tensors from Zivid DeviceArrays

source/Applications/Advanced/Cuda/CaptureAndConvertToDlpackTensorOnCuda
captures a combined 2D+3D frame and builds DLPack tensors directly from
the Zivid DeviceArrays on the CUDA device — both the point cloud
(PointXYZ) and the color image (ColorRGBA_SRGB).

The sample constructs a DLManagedTensor from the building blocks each
DeviceArray exposes (device pointer, shape, strides in elements,
element data type, CUDA device) and validates its fields. The DLPack
data type is derived from the Zivid format's ValueType, so one code
path handles both float point data and 8-bit color. This is the
producer side of DLPack: a consumer such as PyTorch
(torch.from_dlpack), CuPy or JAX imports the resulting tensor for
zero-copy GPU processing. The DeviceArray is kept alive by the
tensor's manager context and released only when the consumer is done
with it, so the GPU memory stays valid as long as the tensor is held.

Building it. The sample is off by default. Configure with
-DUSE_DLPACK=ON in addition to -DUSE_CUDA=ON, and point
DLPACK_INCLUDE_DIR at the directory containing dlpack/dlpack.h
(headers from https://github.com/dmlc/dlpack). CMake fails with a clear
message if the path is missing or wrong.

ZividBenchmark: CSV output

A benchmark run now also writes its results to a CSV file, so a run can
be archived, shared, or compared across machines and cameras instead of
only being read off the console.

• Columns: timestamp, test_category, test_name, iteration,
  median_ms, mean_ms, settings, camera_model, serial_number.
• The settings column records the apertures, exposure times and enabled
  filters used for that test, so rows stay interpretable later.
• System information — API version, OS, camera model, serial number and
  compute device — is written to the file as well.
• The default filename is
  zivid_benchmark_results_<date>_<time>.csv; pass
  --csv-output <filename> to choose your own.

ZividBenchmark: capture-ordering tests are now opt-in

The six capture-ordering tests — 2D then 3D, 3D then 2D, and 3D
including 2D, each with and without the projector — now run only when
you pass --extended. A default run covers connect/disconnect, 3D, 2D,
3D+2D, 2D+3D, copy-data and save, and finishes correspondingly sooner.
Pass --extended to get the previous full set.

Docs

The HALCON compatibility list now also covers Ubuntu 26.04.

[torbsorb]

No changes from the internal version; except improved README.