[SYCL][CUDA] Fixes context release and unnamed context scope

This PR fixes the following bugs:
 1 Some times upon deletion, the deleted context would stay on the CUDA
   context stack and cause subsequent CUDA operations to fail. To fix
   this the context stack is now checked after destruction and if the
   dead context is still there it is removed.
 2 In one location a ScopedContext was created without being assigned to
   a variable and would therefore die immediately. It should now live
   for the entirety of the surrounding scope.

Signed-off-by: Steffen Larsen <[email protected]>

Author: Steffen Larsen

sycl/plugins/cuda/pi_cuda.cpp

@@ -1178,11 +1178,14 @@ pi_result cuda_piContextRelease(pi_context ctxt) {
     CUcontext cuCtxt = ctxt->get();
     CUcontext current = nullptr;
     cuCtxGetCurrent(&current);
-    if(cuCtxt != current)
-    {
-      PI_CHECK_ERROR(cuCtxSetCurrent(cuCtxt));
+    if (cuCtxt != current) {
+      PI_CHECK_ERROR(cuCtxPushCurrent(cuCtxt));
     }
     PI_CHECK_ERROR(cuCtxSynchronize());
+    cuCtxGetCurrent(&current);
+    if (cuCtxt == current) {
+      PI_CHECK_ERROR(cuCtxPopCurrent(&current));
+    }
     return PI_CHECK_ERROR(cuCtxDestroy(cuCtxt));
   } else {
     // Primary context is not destroyed, but released
@@ -1253,6 +1256,7 @@ pi_result cuda_piMemRelease(pi_mem memObj) {
   pi_result ret = PI_SUCCESS;
 
   try {
+
     // Do nothing if there are other references
     if (memObj->decrement_reference_count() > 0) {
       return PI_SUCCESS;
@@ -1263,7 +1267,7 @@ pi_result cuda_piMemRelease(pi_mem memObj) {
 
     if (!memObj->is_sub_buffer()) {
 
-      ScopedContext(uniqueMemObj->get_context());
+      ScopedContext active(uniqueMemObj->get_context());
 
       switch (uniqueMemObj->allocMode_) {
         case _pi_mem::alloc_mode::classic: