Stop backprop when all gradients are None

We need to stop backprop if all gradient variables are None.

(chainer.Variable(numpy.array([1, 2], 'f')) * 1).backward()

It causes TypeError:

TypeError: unsupported operand type(s) for *: 'int' and 'NoneType'

[niboshi]

In order to avoid that, there could be another option:

1. Flag on a variable whether .grad is manually set. If no grad is set, raise an error. If None is set, just do nothing.

Variable already has ._grad_valid flag. Currently it can be False only in Parameter. We could set it to False by default for variables created as outputs of FunctionNode.apply().

[toslunar]

We need to stop backprop if all gradient variables are None.

Be careful with the meaning of “all”. Recall:

1. v.backward() may gather gradients from variables other than v;
2. a function node may have multiple outputs.

The condition “all the output nodes of a function node f are None” does not imply we can ignore f (i.e. continue the loop) in Variable.backward().

In #5709, a related topic is discussed.