Revert use of ACCESSED_BY for managed memory

This reverts commits:
* d49cc0f
* 76c9653
* 11bd103
* e18cf57

While this is beneficial in some cases it causes very important slow
downs when using atomcis in some cases, so reverting this until we can
figure out what's going on.

Author: npmiller

source/adapters/cuda/usm.cpp

@@ -150,27 +150,15 @@ ur_result_t USMDeviceAllocImpl(void **ResultPtr, ur_context_handle_t,
   return UR_RESULT_SUCCESS;
 }
 
-ur_result_t USMSharedAllocImpl(void **ResultPtr, ur_context_handle_t Context,
-                               ur_device_handle_t CommandDevice,
+ur_result_t USMSharedAllocImpl(void **ResultPtr, ur_context_handle_t,
+                               ur_device_handle_t Device,
                                ur_usm_host_mem_flags_t,
                                ur_usm_device_mem_flags_t, size_t Size,
                                uint32_t Alignment) {
   try {
-    ScopedContext Active(CommandDevice);
+    ScopedContext Active(Device);
     UR_CHECK_ERROR(cuMemAllocManaged((CUdeviceptr *)ResultPtr, Size,
                                      CU_MEM_ATTACH_GLOBAL));
-    if (getAttribute(CommandDevice,
-                     CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS)) {
-      UR_CHECK_ERROR(cuMemAdvise((CUdeviceptr)*ResultPtr, Size,
-                                 CU_MEM_ADVISE_SET_ACCESSED_BY,
-                                 CommandDevice->get()));
-    }
-    for (const auto &Dev : Context->getDevices()) {
-      if (getAttribute(Dev, CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS)) {
-        UR_CHECK_ERROR(cuMemAdvise((CUdeviceptr)*ResultPtr, Size,
-                                   CU_MEM_ADVISE_SET_ACCESSED_BY, Dev->get()));
-      }
-    }
   } catch (ur_result_t Err) {
     return Err;
   }