Minor change which fixes some problems with running dill under doctest

[matsjoyce]

Possible fix for #18. Can @JoshRosen see if this solves the test suite problems?

[nchammas]

@JoshRosen @davies Are we still planning to replace cloudpickle with dill? Looks like it's ready for another look.

[davies]

Right now, is there any problem which cloudpickle have? I does not like cloudpickle either, but if we make this change, maybe will meet some unknown situations.

[nchammas]

@davies I think the reasons for switching are (1) dill is more active and better supported, and (2) it handles some cases better than cloudpickle (see #18 and the comments on SPARK-791).

[mmckerns]

This is the status of requests from SPARK: #49
Create any issues for what works for you in cloudpickle that doesn't in dill. They'll at least be on the radar here.

[JoshRosen]

I'm still interested in using Dill in PySpark. I'm sorry for not following up on my earlier Dill issues (such as this one); at the time, I was still in grad school and this fell by the wayside. Now that I'm working on Spark full-time at Databricks, I should be able to follow through on figuring out what's needed from Dill.