Created by: Xreki

目前Paddle框架缺乏有效的机制来保证compile-time和runtime信息的一致性。故，在op_test里面增加compile-time和runtime信息的检查，目前主要检查当输出LoDTensor时，compile-time和runtime的lod_level是否一致。比如sequence_resghape op的InferShape中没有传递lod_level，所以compile-time输出的lod_level是0，但具体实现的Kernel中有给输出Tensor设置了lod，导致了compile-time和runtime的lod_level信息不一致。op单测会出现如下错误：

473: ======================================================================
473: FAIL: test_check_output (test_sequence_reshape.TestSequenceReshape)
473: ----------------------------------------------------------------------
473: Traceback (most recent call last):
473:   File "/paddle/build_paddle/build_docker_manylinux_cuda9.0/python/paddle/fluid/tests/unittests/test_sequence_reshape.py", line 50, in test_check_output
473:     self.check_output()
473:   File "/paddle/build_paddle/build_docker_manylinux_cuda9.0/python/paddle/fluid/tests/unittests/op_test.py", line 916, in check_output
473:     self.check_compile_vs_runtime(fetch_list, outs)
473:   File "/paddle/build_paddle/build_docker_manylinux_cuda9.0/python/paddle/fluid/tests/unittests/op_test.py", line 873, in check_compile_vs_runtime
473:     ")")
473: AssertionError: The lod_level of Output (Out) is different between compile-time and runtime (0 vs 1)

这个PR为OpTest的check_output增加了一个参数check_compile_vs_runtime：

    def check_output(self,
                     atol=1e-5,
                     no_check_set=None,
                     equal_nan=False,
                     check_dygraph=False,
                     inplace_atol=None,
                     check_compile_vs_runtime=True):

我会先将check_compile_vs_runtime的默认值设置成True跑一遍，看看哪些op存在compile-time、runtime的lod_level不一致的问题。

此外：

• LoDTensorArray也是有lod_level的，目前没加入检查。