Fix managed memory incorrectly classified as kDLCUDAHost in DLPack device mapping

_smv_get_dl_device() treated all buffers that are both device- and
host-accessible as kDLCUDAHost. Managed (unified) memory is also both-
accessible, so it was misclassified. CCCL's make_tma_descriptor then
rejected the descriptor with "Device type must be kDLCUDA or
kDLCUDAManaged".

Preserve the is_managed flag already queried via
CU_POINTER_ATTRIBUTE_IS_MANAGED in _query_memory_attrs(), expose it on
Buffer, and use it in _smv_get_dl_device() to return kDLCUDAManaged for
managed memory.

Fixes: https://nvbugspro.nvidia.com/bug/6044342

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: rparolin

cuda_core/cuda/core/_memory/_buffer.pxd

@@ -12,6 +12,7 @@ cdef struct _MemAttrs:
     int device_id
     bint is_device_accessible
     bint is_host_accessible
+    bint is_managed
 
 
 cdef class Buffer:

cuda_core/cuda/core/_memory/_buffer.pyx

@@ -396,6 +396,12 @@ cdef class Buffer:
         _init_mem_attrs(self)
         return self._mem_attrs.is_host_accessible
 
+    @property
+    def is_managed(self) -> bool:
+        """Return True if this buffer is CUDA managed (unified) memory, otherwise False."""
+        _init_mem_attrs(self)
+        return self._mem_attrs.is_managed
+
     @property
     def is_mapped(self) -> bool:
         """Return True if this buffer is mapped into the process via IPC."""
@@ -459,6 +465,7 @@ cdef inline int _query_memory_attrs(
         out.is_host_accessible = True
         out.is_device_accessible = False
         out.device_id = -1
+        out.is_managed = False
     elif (
         is_managed
         or memory_type == cydriver.CUmemorytype.CU_MEMORYTYPE_HOST
@@ -467,10 +474,12 @@ cdef inline int _query_memory_attrs(
         out.is_host_accessible = True
         out.is_device_accessible = True
         out.device_id = device_id
+        out.is_managed = is_managed
     elif memory_type == cydriver.CUmemorytype.CU_MEMORYTYPE_DEVICE:
         out.is_host_accessible = False
         out.is_device_accessible = True
         out.device_id = device_id
+        out.is_managed = False
     else:
         with cython.gil:
             raise ValueError(f"Unsupported memory type: {memory_type}")

cuda_core/cuda/core/_memoryview.pyx

@@ -607,8 +607,10 @@ cdef inline int _smv_get_dl_device(
             device_type = _kDLCUDA
             device_id = buf.device_id
         elif d and h:
-            # We do not currently differentiate pinned vs managed here.
-            device_type = _kDLCUDAHost
+            if buf.is_managed:
+                device_type = _kDLCUDAManaged
+            else:
+                device_type = _kDLCUDAHost
             device_id = 0
         elif (not d) and h:
             device_type = _kDLCPU