[AI] Fix Windows CUDA/OpenVINO support and prevent CUDA driver mismatch crash on Linux

[andriiryzhkov]

Two fixes for AI backend reliability on Windows and Linux:

• Windows CUDA/OpenVINO support: fix EP probing, library detection, and DLL loading so CUDA and OpenVINO providers work on Windows alongside the bundled DirectML build
• Linux CUDA driver check: detect driver/runtime version mismatch before enabling CUDA EP to prevent ORT's abort() crash during inference

Windows changes

• Enable CUDA and OpenVINO provider visibility on Windows (backend_common.c)
• Add Windows DLL search paths for user-installed ORT libraries (%LOCALAPPDATA%/onnxruntime-cuda/ etc.)
• Scan for onnxruntime.dll instead of .so files
• Set SetDllDirectoryW to the custom ORT directory so provider DLLs (CUDA, cuDNN, cublas) are found by LoadLibrary
• Look up EP symbols in the custom-loaded g_ort_module first, fall back to system onnxruntime.dll – fixes the case where the bundled DirectML build doesn't have CUDA symbols

Linux CUDA driver check

• New _check_cuda_driver_compat(): loads libcudart.so, calls cudaDriverGetVersion / cudaRuntimeGetVersion, compares versions
• If driver < runtime, logs a clear message and disables CUDA EP instead of letting ORT abort() mid-inference
• Result is cached per process – check runs only once
• Called from _try_provider before any CUDA EP symbol lookup

[TurboGit]

Thanks!