Add GGML CUDA backend support in InferenceEngine and GGML native bridge

[zhongkaifu]

Motivation

• Enable use of the GGML CUDA backend so GGML-based models can run on NVIDIA GPUs where available.
• Make the managed GGML layer and native bridge detect, initialize and report CUDA support and improve native library resolution across platforms.

Description

• Added BackendType.GgmlCuda and wired model initialization in InferenceEngine.ModelBase to create a GgmlContext with GgmlBackendType.Cuda and use GgmlAllocator for CUDA-backed models.
• Extended TensorSharp.GGML surface: added GgmlBackendType.Cuda, updated GgmlAllocator.Bl​asEnum to return BlasEnum.CUDA for CUDA backends, and improved native library probing to look for platform-specific filenames (.dll, .so, .dylib).
• Updated native bridge (TensorSharp.GGML.Native/ggml_ops.cpp) to detect ggml-cuda.h when present, added BACKEND_TYPE_CUDA, and initialize CUDA with ggml_backend_cuda_init(0) when compiled in, while returning a clear error if CUDA support was not built into the native bridge.
• Made CMakeLists.txt configurable to enable CUDA using TSG_ENABLE_CUDA and gated Metal/ObjC language and flags to Apple builds only so the native build can be targeted per-platform.

Testing

• Attempted dotnet build InferenceEngine/InferenceEngine.csproj -v minimal, but dotnet was not available in the environment so the managed build could not be executed.
• No native cmake/make build was run in this environment, so native bridge compilation with CUDA was not validated by automated builds here.

---

Codex Task