cuda.core: import fails with undefined symbol cudaGetErrorString when _tensor_map is built with -z now (after #2940) #3098

Since #2940 reactivated _tensor_map on CCCL's <cuda/tma>, the extension references the CUDA runtime: cuda/__driver/entry_point.h calls cudaGetErrorString. _tensor_map is not linked against libcudart, and nothing in cuda.core loads cudart before cuda/core/__init__.py imports the module.

With the default flags of a pip build the reference is lazily bound and the import succeeds. With a toolchain that links with -Wl,-z,now, which the conda-forge compilers set in LDFLAGS, the extension carries BIND_NOW and every import cuda.core fails:

ImportError: .../cuda/core/_tensor_map.cpython-314-x86_64-linux-gnu.so: undefined symbol: cudaGetErrorString

Reproduction

Linux, conda-forge environment with Python 3.14, the CUDA 13.4 toolkit, and cuda-bindings 13.4.1; editable build of main.

• readelf -d on the extension shows BIND_NOW and no libcudart entry under NEEDED.
• nm -D --undefined-only lists cudaGetErrorString.
• LD_PRELOAD=<env>/lib/libcudart.so.13 python -c "import cuda.core" succeeds; the plain import fails.

CI does not see this because its wheels are built with lazy binding. The conda-forge feedstock builds with the same compilers as the reproduction, so a conda package built from main would likely fail at import unless cudart is already loaded in the process.

Possible fixes

• Load cudart before _tensor_map is imported, for example cuda.pathfinder.load_nvidia_dynamic_lib("cudart") at the top of _tensor_map.pyx. This also covers the lazy-binding case, where the symbol is first needed at run time.
• Link _tensor_map against cudart in build_hooks.py.

Affects cuda_core/cuda/core/_tensor_map.pyx and cuda_core/build_hooks.py.