Fix bug in graph engine for shared variable

[KazukiYoshiyama-sony]

Two bug fixes:

1. no accumulation of gradients for any shared variable, of which grad.zero is called, into a same function instance
2. no accumulation of a shared and leaf variable into any function instance when using nn.grad

With this bug fixes, the following examples works as expected.

Ex.1

x = nn.Variable.from_numpy_array(np.ones((2, 3))).apply(need_grad=True)
x.grad.zero()
y = x * x * x
y.forward()
y.backward()
print(x.g)  # 3 * x ** 2

Ex. 2

x = nn.Variable.from_numpy_array(np.ones((2, 3))).apply(need_grad=True)
y = x * x * x
x.grad.zero()
dy_dx = nn.grad([y], [x])
dy_dx[0].forward()
print(dy_dx[0].d)  # 3 * x ** 2
dy_dx[0].backward()
print(x.g)  # 6 * x