Created by: Aurelius84

In this PR, IsControlFlowIfVisitor is added to distinguish if/else is plain python statement or control_flow statement.

There are some case should be considered as control flow if/else:

• case 1: if Tensor.numpy()[i] > 2 or if fluid.layers.XXX(x).numpy()[i] > 2
• case 2: if Tensor.shape[i] > 2 where Tensor.shape[i] is -1 or None in compile time.

Currently, we firstly support control flow statement case 1:

• involve Tensor.numpy()[i]

In case 2，Tensor.shape is unknown in gast and should be inferred by other method. Moreover, pred in ConditionalBlock require variable, which means all vars should be Tensor or transformed into Tensor, like fill_constant(shape=[1], dtype='int32', value=Tensor.shape[i]).

** Examples:**

# node is not ast.Compare
node = gast.parse("a + b")
is_control_flow_if(node) == False

# node is not ast.Compare
node = gast.parse("a + x.numpy()[1]")
is_control_flow_if(node) == False

# node is not dependent on Tensor
node = gast.parse("x is None")
is_control_flow_if(node) == False

# `is None` is not dependent on Tensor.data
node = gast.parse("fluid.layers.sum(x) is None")
is_control_flow_if(node) == False

node = gast.parse("fluid.layers.sum(x).numpy() != None")
is_control_flow_if(node) == False

# .numpy() is Dygraph api and return shape=[1] boolean
node = gast.parse("x.numpy()[1] > 1")
is_control_flow_if(node) == True

# complex statement, but include control flow `if` in sub_expr
node = gast.parse("x is not None and 1 < x.numpy()[1]")
is_control_flow_if(node) == True

node = gast.parse("1 < fluid.layers.sum(x).numpy()[2] or x+y < 0")
is_control_flow_if(node) == True

Advanced features will be implemented while we test in real models.