[CUDA][Bindless] Fix memory leak in interop mapping

* added a map to ur_device_handle_t_.
* Capture the leaking Cumipmappedarray in
  urBindlessImagesMapExternalArrayExp into map.
* Update urBindlessImagesImageFreeExp to check if the Cuarray is
  derived, then destroy the corresponding Cumipmappedarray.

Author: cppchedy

source/adapters/cuda/device.hpp

@@ -114,6 +114,9 @@ struct ur_device_handle_t_ {
   bool maxLocalMemSizeChosen() { return MaxLocalMemSizeChosen; };
 
   uint32_t getNumComputeUnits() const noexcept { return NumComputeUnits; };
+
+  // bookkeeping for mipmappedArray leaks in Mapping external Memory
+  std::map<CUarray, CUmipmappedArray> ChildCuarrayFromMipmapMap;
 };
 
 int getAttribute(ur_device_handle_t Device, CUdevice_attribute Attribute);

source/adapters/cuda/image.cpp

@@ -430,6 +430,11 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageFreeExp(
   ScopedContext Active(hDevice);
   try {
     UR_CHECK_ERROR(cuArrayDestroy((CUarray)hImageMem));
+    if (auto it = hDevice->ChildCuarrayFromMipmapMap.find((CUarray)hImageMem);
+        it != hDevice->ChildCuarrayFromMipmapMap.end()) {
+      UR_CHECK_ERROR(cuMipmappedArrayDestroy((CUmipmappedArray)it->second));
+      hDevice->ChildCuarrayFromMipmapMap.erase(it);
+    }
   } catch (ur_result_t Err) {
     return Err;
   } catch (...) {
@@ -1104,6 +1109,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesMapExternalArrayExp(
       CUarray memArray;
       UR_CHECK_ERROR(cuMipmappedArrayGetLevel(&memArray, memMipMap, 0));
 
+      hDevice->ChildCuarrayFromMipmapMap.emplace(memArray, memMipMap);
+
       *phImageMem = (ur_exp_image_mem_native_handle_t)memArray;
     }