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)
```

From: @yanglf1121
Reviewed-by: @liangchenghui
Signed-off-by:

From: @liangzhibo
Reviewed-by: @ginfung,@zh_qh
Signed-off-by: @zh_qh

From: @yuan_shen_zhou
Reviewed-by: @liangchenghui
Signed-off-by: @liangchenghui

Merge pull request !3940 from zhangbuxue/support_interface_all_and_any_of_tensor

* add infer_value for common math ops
* convert constant bool tensor to bool value

* do not infer value when encounter 0 as division

for while condition, do not unrool if condition is a tensor

Merge pull request !1803 from Kang/master

support tensor set value by tensor index

Merge pull request !700 from penn/validate_bprop_rules

Signed-off-by: candanzg <zhangshucheng@huawei.com>

1) A[B]=U
2) A[A>n]=U
    A.shape == B.shape
    U is a scalar or Tensor(size==1)
    B is Tensor(dtype=bool)
    n is a Number

Signed-off-by: candanzg <zhangshucheng@huawei.com>