From: @wang_hua_2019
Reviewed-by: @c_34
Signed-off-by:

The basic idea is: exploits data dependency to control the execution order
of side-effect operations, and keep the semantics of ANF unchanged.

The ControlDepend primitive is removed and there are two primitives added:

1. UpdateState:
```
  a = Assign(para, value)
```
became:
```
  a = Assign(para, value, u)
  u = UpdateState(u, a)
```

2. Load:
```
  x = Add(para, value)
```
became:
```
  p = Load(para, u)
  x = Add(p, value)
  u = UpdateState(u, p)
```

Merge pull request !6018 from h.farahat/ms_nonsink_epochs

Merge pull request !6724 from TuDouNi/master

Merge pull request !5482 from liuyang/md_save_checkpoint

Merge pull request !5351 from yao_yf/parallel_context_collation

Merge pull request !4814 from dengyutao/tinybert

Merge pull request !4187 from hanhuifeng/tinybert_suit_gpu