#1358 build: vmm.cpp builds with a CUDA toolkit older than 12.5 again (#1071)

本文

Fixes #1071.

`src/core/vmm.cpp` (new in 0.1.40) calls `cudaGetDriverEntryPointByVersion`, which arrived in CUDA 12.5, so the
engine no longer builds with an older toolkit. This takes the same `CUDART_VERSION` guard that
`src/core/expert_cache.cpp` already uses for the same call.

How I hit it: Ubuntu 24.04, 2x RTX 4090, driver 595.84. `./update.sh` from 0.1.39 found the distribution's nvcc
(`nvidia-cuda-toolkit` 12.0.140, which setup accepts for sm_89) and stopped at `vmm.cpp:28` with
`'cudaGetDriverEntryPointByVersion' was not declared in this scope`. Setup then kept engine 0.1.39, so the server was
0.1.40.1 and the engine 0.1.39.

Checked:
- With the patch the engine builds with that CUDA 12.0.140 nvcc (sm_89, host GCC 12.4).
- On CUDA 12.5 or newer the guard takes today's call, so CUDA 13 builds compile the same code as before.
- `tools/test_setup_*.py`: all 23 pass (Linux).