Created by: dzhwinter

This PR only contains single machine version, some conflict happens when distributed training are enabled. It contains 2 main modifies.

1. Python memory optimize equally implement in c++ IR graph. by default, it is turned off, can be triggered with "memory_optimize" in pybind;
2. Implement a early_delete_pass. can be triggered with "memory_early_delete" in pybind.

model tests for QA https://github.com/dzhwinter/modeltests

Following PR should be contained in this one, but it too huge and will takes more time to pass all the tests.

• add memory optimize for executor(now only in parallelexecutor).
• c++ inference use IR memory optimize, add memory optimize pass in inference passes
• test performance in inference cases.
• fix confilict with distributed training jobs. Especially in merge_multi_batch_pass enabled. [doing]
• add inplace pass before c++ IR memory optimize pass, to save more memory.