Fix CUDA minimal build requiring CUDNN_HOME (#27308)
When building with `onnxruntime_CUDA_MINIMAL=ON`, cuDNN is not needed
since minimal CUDA
builds only use basic CUDA runtime (e.g., for TensorRT-only builds).
Skip `cuDNN.cmake`
inclusion and `CUDNN_INCLUDE_DIR` references when `CUDA_MINIMAL` is
enabled.
Fixes #24361
### Description
Make cuDNN dependency conditional on `NOT onnxruntime_CUDA_MINIMAL`:
- `cmake/external/onnxruntime_external_deps.cmake`: Skip
`include(cuDNN)`
- `cmake/onnxruntime_python.cmake`: Skip `CUDNN_INCLUDE_DIR` in include
paths
- `cmake/onnxruntime_unittests.cmake`: Skip `CUDNN_INCLUDE_DIR` in two
places
### Motivation and Context
When building with `onnxruntime_CUDA_MINIMAL=ON` (e.g., TensorRT-only
builds), the build
fails with:
CMake Error at external/cuDNN.cmake:3 (find_path):
Could not find CUDNN_INCLUDE_DIR using the following files: cudnn.h
cuDNN is not needed for minimal CUDA builds since TensorRT has its own
optimized kernels.
This fix allows building without setting `CUDNN_HOME`.
