Add timedeltas strategy to datetime extra

[Daenyth]

It seems appropriate.

[Zac-HD]

@malinoff, @NotBobTheBuilder, do either of you want to finish up your pulls? If not, I can try integrating the implementation from #441 with the tests and docs from #352 so you both get appropriate credit.

[NotBobTheBuilder]

Sure - what needs finishing on mine?

[Zac-HD]

I've basically outlined it in #520, where I pulled in some tests and extra docs from @malinoff's implementation.

• git pull https://github.com/Zac-HD/hypothesis-python.git timedelta-strategies to get the merged version. I've inlined the validation logic, to avoid adding new functions to the namespace which resemble incompatible things, and fixed a few linting issues.
• Remove test_can_find_zero_delta, as it's a poor duplicate of test_simplifies_towards_zero_delta
• Change the generation strategy so that timedeltas shrink towards 0, preferring positive values. If they already do, check that the tests are working (see my travis build). Note in the documentation that timedelta normalisation means this will give some unintuitive results (eg timedelta(microseconds=-1) looks pretty scary).
• Ensure that the tests have complete code coverage; this usually just means passing invalid inputs so that all the validation logic gets exercised. If it's anything more complicated, I'd be happy to help out.