Created by: kuke

This pull request tries to define the way to implement a complex number neural network in PaddlePaddle.

Major changes and features:

• Add a new class ComplexVariable to hold complex numbers, and expose frequently used attributes name, shape and dtype for it;
• Support methods to_variable(), .numpy() and print() to make sure that ComplexVariable has the same interaction experience as the Variable defined on real number.
• Users can also visit the members and methods of the real or imaginary part of ComplexVariable seperately as a common Variable.
• Lay down API rules. Users only need to change the prefix of APIs from paddle to paddle.complex to build a complex number model, as the new added layers for element-wise complex number add/sub/mul/div demonstrated.

The usage (only support DyGraph at present):

>>> import numpy as np
>>> import paddle
>>> import paddle.fluid.dygraph as dg
>>> a = np.array([[1.0+1.0j, 2.0+1.0j], [3.0+1.0j, 4.0+1.0j]])
>>> b = np.array([[1.0+1.0j, 1.0+1.0j]])
>>> c = np.array([[1.0, 2.0], [3.0, 4.0]])
>>> with dg.guard():
...     x = dg.to_variable(a, "complex_x")
...     y = dg.to_variable(b)
...     z = dg.to_variable(c)
...     out = paddle.complex.elementwise_add(x, y)
...     print(out)
...
REAL: name tmp_1, dtype: VarType.FP64 shape: [2L, 2L] 	lod: {}
	dim: 2, 2
	layout: NCHW
	dtype: double
	data: [2 3 4 5]
IMAG: name tmp_2, dtype: VarType.FP64 shape: [2L, 2L] 	lod: {}
	dim: 2, 2
	layout: NCHW
	dtype: double
	data: [2 2 2 2]

>>> print(type(z), type(out))
(<class 'paddle.fluid.core_avx.VarBase'>, <class 'paddle.fluid.framework.ComplexVariable'>)
>>> print(out.numpy())
[[2.+2.j 3.+2.j]
 [4.+2.j 5.+2.j]]
>>> out.name
{'real': u'tmp_1', 'imag': u'tmp_2'}
>>> out.name = "out"
>>> out.name
{'real': u'out.real', 'imag': u'out.imag'}
>>> out.real.name = "new_name"
>>> out.name
{'real': u'new_name', 'imag': u'out.imag'}
>>> out.dtype
'complex128'
>>> out.shape
[2L, 2L]
>>> out.real.stop_gradient = True
>>> out.imag.stop_gradient = False
>>> out.real.stop_gradient
True
>>> out.imag.stop_gradient
False