* Fixed allclose_op bug, which cannot deal with some cases of fp64 inputs.

* improved CUDA kernel performance.

* Fixed a bug in cuda kernel which cannot deal with large dimension input, and added an unit test for it.

* Add a test case for float32 input.