Created by: liym27

1. Support to create LoDTensorArray in control flow (cond and while_loop)

For example:

# Create a null array in control flow sub-block
def true_fn():
    null_array = create_array("float32")
    # do something else
    return

def false_fn():
   # do something else
   return

res = cond(A==B, true_fn, false_fn)

Before this PR, it throws an error:

ValueError: Var array_0.out is not found recursively

Because null_array(array_0.out ) is considered as input not output of OP.

2. Fix bug: return LoDTensorArray in while_loop

Before this PR, while_loop doesn't assign LoDTensorArray to parent block. It causes that the returned LoDTensorArray can't be created in while block, which should assign to parent block. For example:

def cond(i, array_parent): 
    return i < ten

def body(i, array_parent): 
    i = i + 1
    x = fluid.layers.ones(shape=[3], dtype="int32")
    array_sub = fluid.layers.create_array("int32")
    for j in range(5):
        x = func(array_parent[j])
        tensor_j = fluid.layers.fill_constant(shape=[1], dtype="int64", value=j)
        fluid.layers.array_write(x,tensor_j, array_sub)
    array_parent = array_sub
    return i, array_parent

# do something

i = layers.fill_constant(shape=[1], dtype='int64', value=0)
ten = layers.fill_constant(shape=[1], dtype='int64', value=10)
i, array_parent = layers.while_loop(cond, body, [i, array_parent])

3. Delete unnecessary code in list_transformer.py