Holoscan Camera Python support

Holoscan Camera exposes native camera capture operators to Holoscan SDK 5 Python graphs. Capture, buffer ownership, and device lifecycle remain in C++. Python applications use the SDK's typed image payloads and tensor APIs.

Import operators from holoscan_camera.operators. The SDK also discovers the package through the holoscan.camera alias. A Python operator facade loads its backend and resolves a Core registry proxy; it does not implement capture. Wheel builds configure facades and typing declarations for all three backends. Usability depends on the wheel's native extensions, host hardware, and runtime libraries. An advertised name or .pyi declaration is not a capability check. SIPLCaptureService and its related value types are bound directly by _sipl.

The default wheel build includes V4L2. Running capture requires a compatible camera, the matching Holoscan Core CUDA 13 package, and the platform's native runtime libraries. Wheels do not bundle camera drivers or the system CUDA runtime. Install the artifact qualified for your target platform.

Source builds require CMake 4+, a C++20 toolchain, CPython 3.12+, and a matching Holoscan SDK development installation with Python support and its exported Python build helpers. Build and application instructions are provided in the source distribution's python/README.md and application directories.

The ARM64 wheel includes V4L2 and separate R39 SIPL v2 and Argus extensions. V4L2 remains usable on SBSA/DGX Spark and R38 without loading those backends. R39 SIPL requires supported Orin/Thor hardware; R39 Argus requires Orin. The x86_64 wheel supports V4L2. Unsupported backend imports may fail; wildcard imports request all backends. Attribute probes such as hasattr() and inspect.getmembers() can also request a backend and propagate ImportError. R38 SIPL is not supported by the wheel.