Add an attribute to FunctionNode to indicate it supports arrays on different devices

[niboshi]

This idea came from an offline discussion related to #454.

Currently only array types are checked in FunctionNode inputs, but not their devices.
If the arrays are on different GPU devices, it causes an error that comes from CuPy, not easy to investigate the fault. It's better for users if Chainer can catch such error earlier and present more easy-to-understand error message.

On the other hand, there may exist (user-defined) functions that support arrays on multiple GPU devices.

The proposed solution is:

• Introduce an attribute in FunctionNode to indicate whether it supports multiple devices. It's "disabled" by default (i.e. only single device is supported).
• If it's "disabled", Chainer checks array devices in function inputs.
• If a FunctionNode implementation wants to support multiple device, it should be explicitly set to "enabled".

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed after 30 days if no further activity occurs. Thank you for your contributions.

[kmaehashi]

This should solve issues like #4144 (GPU mismatch).