#1359 setup: link the engine against the toolkit whose nvcc builds it (CUDAToolkit_ROOT)

描述

When a second CUDA toolkit sits next to the distribution's one, CMake compiles with the chosen nvcc but can take
cudart and cuBLAS from the other toolkit.

Measured on Ubuntu 24.04 with the distribution's CUDA 12.0 (`nvidia-cuda-toolkit`, in /usr) and CUDA 13.0.2 in
/opt/cuda-13.0 (NVIDIA's redist archives), `STRATA_NVCC=/opt/cuda-13.0/bin/nvcc`, cmake 4.4.3 from `.venv`:

| | `CUDA_cudart_LIBRARY` in CMakeCache | engine's NEEDED |
|---|---|---|
| before | /usr/lib/x86_64-linux-gnu/libcudart.so | libcudart.so.12, libcublas.so.12 (with the 13.0 headers) |
| after | /opt/cuda-13.0/lib/libcudart.so | libcudart.so.13, libcublas.so.13 |

The change passes `-DCUDAToolkit_ROOT=<nvcc>/../..` to the engine build and to the CUDA image-encoder build. For the
distribution's own `/usr/bin/nvcc` nothing is added, so that case keeps CMake's search as before. A build folder
configured earlier keeps its cached library paths; a fresh configure picks the right ones.

An engine built this way (0.1.40, CUDA 13.0.88, the same root given through the environment) serves UD-IQ4_XS on a
2x RTX 4090 split here.

Tests: all 23 `tools/test_setup_*.py` pass (Linux).