StarPU: refactor GPU codelets

Change the allocation of the cl_args to have a malloc instead of stack allocation to make sure the GPU kernels are referencing something still allocated when submitting the kernels.

runtime/starpu/codelets/codelet_zcesca.c

@@ -46,7 +46,6 @@ static void cl_zcesca_cpu_func(void *descr[], void *cl_arg)
     Dj = cti_interface_get(descr[4]);
     A  = cti_interface_get(descr[5]);
 
-    starpu_codelet_unpack_args( cl_arg, &clargs );
     TCORE_zcesca( clargs->center, clargs->scale, clargs->axis,
                   clargs->m, clargs->n, clargs->mt, clargs->nt,
                   Gi, Gj, G, Di, Dj, A );