Fix managed memory DLPack device type on buffer-side export paths

Update setup_dl_tensor_device() and Buffer.__dlpack_device__() to emit
kDLCUDAManaged for managed memory, closing the gap where the Buffer ->
DLPack capsule -> StridedMemoryView path still misclassified managed
buffers as kDLCUDAHost. Add cross-reference comments to keep the three
classification sites aligned.

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

Author: rparolin

cuda_core/cuda/core/_dlpack.pyx

@@ -95,7 +95,8 @@ cdef inline int setup_dl_tensor_device(DLTensor* dl_tensor, object buf) except -
         device.device_type = _kDLCUDA
         device.device_id = buf.device_id
     elif buf.is_device_accessible and buf.is_host_accessible:
-        device.device_type = _kDLCUDAHost
+        # Keep in sync with Buffer.__dlpack_device__() and _smv_get_dl_device().
+        device.device_type = _kDLCUDAManaged if buf.is_managed else _kDLCUDAHost
         device.device_id = 0
     elif not buf.is_device_accessible and buf.is_host_accessible:
         device.device_type = _kDLCPU

cuda_core/cuda/core/_memory/_buffer.pyx

@@ -328,7 +328,9 @@ cdef class Buffer:
         if d and (not h):
             return (DLDeviceType.kDLCUDA, self.device_id)
         if d and h:
-            # TODO: this can also be kDLCUDAManaged, we need more fine-grained checks
+            # Keep in sync with setup_dl_tensor_device() and _smv_get_dl_device().
+            if self.is_managed:
+                return (DLDeviceType.kDLCUDAManaged, 0)
             return (DLDeviceType.kDLCUDAHost, 0)
         if (not d) and h:
             return (DLDeviceType.kDLCPU, 0)

cuda_core/cuda/core/_memoryview.pyx

@@ -607,6 +607,7 @@ cdef inline int _smv_get_dl_device(
             device_type = _kDLCUDA
             device_id = buf.device_id
         elif d and h:
+            # Keep in sync with Buffer.__dlpack_device__() and setup_dl_tensor_device().
             if buf.is_managed:
                 device_type = _kDLCUDAManaged
             else: