[AI] GPU-accelerated ORT install scripts and documentation

[andriiryzhkov]

Follow-up to #20647 (code changes). This PR adds the install helper scripts and documentation.

Helper scripts for Linux and Windows users to install ONNX Runtime with GPU acceleration, used together with the plugins/ai/ort_library_path preference or DT_ORT_LIBRARY env var added in #20647.

Scripts

Script	GPU	Method	Time
install-ort-nvidia.sh	NVIDIA	Download prebuilt from GitHub	~30 sec
install-ort-nvidia.ps1	NVIDIA (Windows)	Download prebuilt from GitHub	~30 sec
install-ort-amd.sh	AMD	Download prebuilt wheel from AMD repo	~30 sec
install-ort-amd-build.sh	AMD	Build from source with MIGraphX EP	10–20 min
install-ort-intel.sh	Intel	Download prebuilt wheel from PyPI	~30 sec

• Auto-detect system (distro, driver version, ROCm version)
• Validate prerequisites before downloading/building
• Distro-specific install hints for missing dependencies (Ubuntu, Fedora, Arch, openSUSE)
• AMD build script matches ORT version to installed ROCm (6.0–7.2 compatibility matrix)
• AMD prebuilt script extracts the shared library from the pip wheel (no Python required)

Documentation

tools/ai/README.md covers:

• What's bundled by default per platform (CPU/DirectML/CoreML)
• Requirements per GPU vendor
• Manual install instructions for users who prefer not to use the scripts

macOS

No scripts needed – CoreML is bundled and statically linked.