Created by: chengduoZH

Dependent PR: #14208, #14202, #14201

This PR adds switch_to_fp16 to allow the user to specify some operators's computation with fp16. For example:

        import paddle.fluid as fluid
        img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32')
        label = fluid.layers.data(name='label', shape=[1], dtype='int64')
        with fluid.contrib.switch_to_fp16(fluid.default_main_program()):
            prediction = fluid.layers.fc(input=img, size=200, act='tanh')
        prediction = fluid.layers.cast(prediction, np.float32)
        loss = fluid.layers.cross_entropy(input=prediction, label=label)

switch_to_fp16 casts the dtype of the operator's input to fp16, and the dtype of the output is inferred by calling infer_var_type, in most case, the output's dtype is fp16 too.

Note: some operator should not apply fp16 computation, because the output of those operators maybe overflows. For example, exp, log, softmax, these operators' output's range is larger then it's input's.

In order to guide users to use fp16, this PR adds white_list, black_list and gray_list.

• white_list is the collection of operators which is safe when using fp16 computation.
• black_list is the collection of operators which may be overflows when using fp16 computation.
• gray_list is the collection of operators whose gradient may be overflows when using fp16 computation.

These lists are incomplete now, please feel free to make it completed.