Created by: Aurelius84

Adds ProgramCache

In this PR, FunctionCache and ProgramCache are added to avoid transforming code of dygrah function by AST analyze and inserting ops into main_program, respectively.

That means program will be built once, even users put myNet(input) into for loop.

code example:

def train_mnist():
    with fluid.program_guard(fluid.Program()):
        mnist = MNIST()

        for epoch in range(self.epoch_num):
            for batch_id, data in enumerate(self.train_reader()):
                dy_x_data = np.array([
                    x[0].reshape(1, 28, 28) for x in data
                ]).astype('float32')
                y_data = np.array([x[1] for x in data
                                    ]).astype('int64').reshape(-1, 1)
                # fetch outputs dynamically as dygraph model.
                # Program will be built only once.
                prediction, acc, avg_loss = mnist(dy_x_data, label=y_data)

Moreover, code related to exe.run is moved into ProgramCache to allow users specifying optimizer which used to minimize loss.