fix RDD.reduce when rdd contains empty partitions

[tools4origins]

Fixes #83

It let the TypeError to be thrown instead of checking the emptiness of the partition beforehand as values is a generator and it seems better not to affect it.

[svenkreiss]

Thanks for submitting this and sorry for the long delay!

I have a concern about the nested reducer. I think (and I haven't tested it) it is not importable which creates problems for the built-in pickle. Generally, pysparkling must work without cloudpickle, but it turns out there was no unit test for it. #86 adds that. I can probably merge that rather quickly. Do you mind testing your branch with this additional unit test? I think it will require a small change to make the reducer importable.

[svenkreiss]

Actually, I'll do it the other way around. Your PR is good for the current set of tests. So I will merge that now and see how to make #86 work afterwards.

[svenkreiss]

And I was wrong. Your change is not a problem for the built-in pickle.
Thanks a lot for your contribution!