fix bugs under kSgdSparseCpuTraining mode (#100)

Local training with "sparse_update = True" parameter triggers kSgdSparseCpuTraining mode, fix bugs under it.

paddle/trainer/ThreadParameterUpdater.cpp

@@ -141,7 +141,7 @@ void SgdThreadUpdater::traverse(GetTraverseCallback getTraverseCallback) {
   } else if (hasCpuPara) {
     getGlobalSyncThreadPool()->exec(cpuTraverse);
   } else if (hasGpuPara) {
-    cpuTraverse(0, 0);
+      gpuTraverse(0, 0);
   }
 }
 

paddle/trainer/TrainerInternal.cpp

@@ -101,6 +101,7 @@ void TrainerInternal::trainOneBatch(int64_t batchId,
       // it
       //! to ParameterHook.
       auto& grad = para->getBuf(PARAMETER_GRADIENT);
+      SetDevice device(para->getDeviceId());
       paraStats[para->getID()].avgAbsGrad = grad->getAbsSum() / para->getSize();
       paraStats[para->getID()].maxAbsGrad = grad->getAbsMax();
     }