Correct tensor copy misuse

[JiayiFeng]

Correct misusing of TensorCopy in following ops or op tests:

concat_test
sequence_padding_test
multiplex_op
sequence_slice_op
lod_reset_op

[chengduoZH]

If we replace cudaMalloc with cudaMallocHost, the memory::Copy should have sync and async operation respectively, because someplace call memory::Copy directly. For example

Paddle/paddle/fluid/framework/tensor_util.h

Lines 104 to 105 in 99acf1d

	memory::Copy(dst_place, dst_ptr,
	boost::get<platform::CPUPlace>(src.place()), src_ptr, size);

[JiayiFeng]

In current design, memcpy takes an cudaStream_t argument. If this argument is nullptr, the memcpy will be synchronous. @chengduoZH

[chengduoZH]

LGTM!