Fix MSVC CCCL preprocessor issue and add comment about CCCL fix

Author: seberg

cuda_core/build_hooks.py

@@ -245,7 +245,8 @@ def get_sources(mod_name):
     extra_link_args = []
     extra_cythonize_kwargs = {}
     if sys.platform == "win32":
-        extra_compile_args += ["/std:c++17"]
+        # CCCL headers #error without the conforming MSVC preprocessor.
+        extra_compile_args += ["/std:c++17", "/Zc:preprocessor"]
         if debug:
             raise RuntimeError("Debuggable builds are not supported on Windows.")
     else:

cuda_core/cuda/core/_tensor_map.pyx

@@ -686,6 +686,8 @@ cdef class TensorMapDescriptor:
         # TODO(seberg): CCCL has various limitations as of writing, see:
         # https://github.com/NVIDIA/cccl/issues/10511.
         # Because of this, the path is only taken in _very_ narrow conditions.
+        # We should probably use it broadly as soon as the fix is available in
+        # CCCL 3.5.0 (and likely just not bother with these narrow conditions).
         make_tma = _get_cccl_make_tma_fn()
         if (make_tma != NULL
                 and device_type == _kDLCUDA

cuda_core/cuda/core/_tensor_map_cccl.pyi

@@ -1,4 +1,4 @@
-# This file was generated by stubgen-pyx v0.2.6 from cuda_core/cuda/core/_tensor_map_cccl.pyx
+# This file was generated by stubgen-pyx v0.2.19 from cuda_core/cuda/core/_tensor_map_cccl.pyx
 
 """Optional CCCL <cuda/tma> helper for TensorMapDescriptor.
 
@@ -9,4 +9,3 @@ shared cudart. It must not be imported at ``cuda.core`` package import time.
 ``get_make_tma_descriptor_tiled`` is exported via ``__pyx_capi__`` for
 soft-linking from ``_tensor_map.pyx`` (returns real fn or NULL).
 """
-from __future__ import annotations