add some synchro in cuda gelqt, tslqt and tsqrt kernel to improve algo robustness

cudablas/compute/cuda_zgelqt.c

@@ -142,6 +142,7 @@ int CUDA_zgelqt(
       magma_zlarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaRowwise,
                         rows, cols, ib, da_ref(i,i), ldda, dt_ref(0,i),
                         lddt, da_ref(i+ib,i), ldda, dwork, lddwork);
+      cudaThreadSynchronize();
       old_i = i;
       old_ib = ib;
       if (i+nb >= k){

cudablas/compute/cuda_zgemerge.c

@@ -115,7 +115,6 @@ int CUDA_zgemerge(
         for(i=0; i<N; i++){
             cola = A + i*LDA;
             colb = B + i*LDB;
-//            cublasZcopy(i+1, cola, 1, colb, 1);
             cudaMemcpyAsync(colb , cola,
                             (i+1)*sizeof(cuDoubleComplex),
                             cudaMemcpyDeviceToDevice, stream);
@@ -124,7 +123,6 @@ int CUDA_zgemerge(
         for(i=0; i<N; i++){
             cola = A + i*LDA;
             colb = B + i*LDB;
-//            cublasZcopy(M-i, cola + i, 1, colb + i, 1);
             cudaMemcpyAsync(colb+i , cola+i,
                             (M-i)*sizeof(cuDoubleComplex),
                             cudaMemcpyDeviceToDevice, stream);

cudablas/compute/cuda_ztslqt.c

@@ -163,6 +163,7 @@ int CUDA_ztslqt(
                     dwork, lddwork,
                     dwork + nb * lddwork, nb,
                     stream );
+            cudaThreadSynchronize();
             old_i = i;
             old_ib = ib;
         }

cudablas/compute/cuda_ztsqrt.c

@@ -185,6 +185,7 @@ int CUDA_ztsqrt(
                     dwork, ib,
                     dwork + ib * cols, rows,
                     stream );
+            cudaThreadSynchronize();
             old_i = i;
             old_ib = ib;
         }