Add an option that enables dropout behave deterministically to RSGCN.

[delta2323]

See #146 for detailed discussion.

[corochann]

Do we really need deteministic behavior only for testing? If it is ok for current test, I would like to close the issue.

[delta2323]

Do we really need deteministic behavior only for testing?

If we can test the case in which the dropout option is not 0 without deterministic behavior, we do not have to implement it, until some users want it (e.g. a test that checks the dropout function is surely called when dropout ratio is greater than 0). Otherwise, it is better to do it, I think.