Fix installed static MLX package on Windows

[yingjiacai]

Proposed changes

Fix the installed static MLX package on Windows so downstream CMake
projects can link against mlx.lib without requiring mlx.dll.

The changes:

• Compile the dlfcn-win32 implementation directly into the static MLX
  library instead of exposing a FetchContent-only target to installed
  consumers.
• Link the required Windows Psapi.lib system library.
• Represent the downloaded OpenBLAS import library as an imported CMake
  target and install the .lib file.
• Keep cudnn_frontend as a build-only dependency.
• Reconstruct the required CUDA, cuDNN, and OpenBLAS dependencies in the
  installed MLX CMake package.

Testing

Manually verified on Windows using MSVC, Ninja, CUDA 13.0, and cuDNN 9.19:

• Configured and built MLX with BUILD_SHARED_LIBS=OFF
• Installed MLX successfully
• Verified that mlx.lib is installed without requiring mlx.dll
• Rebuilt and ran the downstream mlx_cpp_inference project against the
  installed static MLX package
• Ran pre-commit run --all-files successfully

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)