[CUDA][HIP] Improve command-buffer sync points

Several improvements to sync-point implementation
in HIP and CUDA command-buffer adapters with
additional CTS coverage to back it up.

* In the CUDA/HIP adapters we assume that there is always
  a return sync-point passed by the user. However, this is not
  required by the UR API, so we should check that
  the return value is non-null before dereferencing.
* The Fill helper function is can implement as fill as several commands
  for certain pattern sizes, we were creating a sync point for every
  internal command. This is not required, these commands from a linear
  dependency chain, so only the leaf command is required to be a sync
  point for future commands to depend on.
* Remove `shared_ptr` from `CUgraphNode` objects stored for sync-points.
  `CUgraphNode` is a pointer type, and is managed by the CUDA driver
  runtime rather than us.
* Simplify handling of return results. We don't always use the helper
  macro for returning the `ur_result_t` value no a function call fail,
  and also often unnecessarily use a variable to store return code.
* Use `hipMemcpyDefault` for USM memcopy

Author: Ewan Crawford

source/adapters/cuda/command_buffer.cpp

@@ -699,6 +699,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendUSMPrefetchExp(
     uint32_t numSyncPointsInWaitList,
     const ur_exp_command_buffer_sync_point_t *pSyncPointWaitList,
     ur_exp_command_buffer_sync_point_t *pSyncPoint) {
+  ur_result_t Result = UR_RESULT_SUCCESS;
   // Prefetch cmd is not supported by Cuda Graph.
   // We implement it as an empty node to enforce dependencies.
   CUgraphNode GraphNode;