fix(cuda/msvc): add CUDA 13.4 CUB fixes, Win32 CreateFileW, and graceful tensor name fallback

[vrwallace]

Summary of Changes

1. CUDA 13.4 CUB Header Fixes:

   • Resolved compilation errors with NVIDIA CUDA Toolkit 13.4 (CCCL 3.4) in ggml/src/ggml-cuda/argsort.cu and ggml/src/ggml-cuda/top-k.cu.
   • Restricted version check conditions to safely fall back to native CUDA kernels on CUDA 13.4.

2. Windows MSVC httplib Compatibility:

   • Replaced CreateFile2 with standard Win32 CreateFileW in vendor/cpp-httplib/httplib.cpp to fix MSVC desktop build linking.

3. Graceful Model Tensor Name Fallback:

   • Updated src/llama-arch.cpp to return "unknown" instead of GGML_ABORT() on unrecognized tensor names, preventing process crashes on novel model architectures (e.g. Gemma 4).

Verification

• Built and verified with Visual Studio 2022 (MSVC 19.44) + CUDA Toolkit 13.4 on NVIDIA RTX 4070 Ti.