optimize executor.run

On k40 with 4 devices, time reduces from ~4.0 to ~3.8+, should be
more obvious on better hardware

The error is:

paddle/fluid/operators/math/concat.cc:47:72: error: invalid initialization of non-const reference of type 'paddle::platform::CPUPlace&' from an rvalue of type 'paddle::platform::CPUPlace'
   auto& cpu_place = boost::get<platform::CPUPlace>(context.GetPlace());

Should not use reference for cpu_place.

Single device on se-resnet reduce from 0.56 to 0.50