Fix: Drop -lcuda from stringzillas-cuda wheel link line

The wheel build at `setup.py:412` linked the CUDA driver stub `-lcuda`
in addition to the runtime `-lcudart`, but the library only ever calls
runtime APIs (`cudaMallocManaged`, `cudaFree`, `cudaMemcpyToSymbolAsync`,
`cudaSetDevice`, `cudaGetDeviceProperties`). No driver-API call sites
exist in `include/stringzillas/types.cuh` or `include/stringzillas/similarities.cuh`.

The CMake path is already correct: `CMakeLists.txt:766` links only
`CUDA::cudart` to `stringzillas_cuda_shared`. The wheel path regressed
in 4abf63c (Aug 2025, "Make: Custom CudaBuildExtension for Python")
which added the unconditional `-lcuda` to `extra_link_args` even though
the manylinux wheel containers intentionally omit `libcuda.so` (it's
the driver stub, not part of the CUDA toolkit), causing
`ld: cannot find -lcuda` and breaking every `Build StringZillas-CUDA`
matrix job.

Author: ashvardanian

setup.py

@@ -409,7 +409,7 @@ def windows_settings(use_cpp: bool = False) -> Tuple[List[str], List[str], List[
             ["python/stringzillas.c", "c/stringzillas.cu"],
             include_dirs=["include", "c", "fork_union/include", "/usr/local/cuda/include"],
             extra_compile_args=compile_args,
-            extra_link_args=link_args + ["-L/usr/local/cuda/lib64", "-lcudart", "-lcuda", "-lstdc++"],
+            extra_link_args=link_args + ["-L/usr/local/cuda/lib64", "-lcudart", "-lstdc++"],
             define_macros=[("SZ_DYNAMIC_DISPATCH", "1"), ("SZ_USE_CUDA", "1")] + macros_args,
             language="c++",  # Force C++ linking
         ),