Implement rand_tangent and difference

[willtebbutt]

This PR lays some foundations for (hopefully) moving FiniteDifferences over to ChainRules types, and contains exactly two things:

1. Moves rand_tangent over from ChainRulesTestUtils. It turns out that it's useful here, and since ChainRulesTestUtils depends on this package, there shouldn't be any harm implementing them here.
2. Implements difference, which is more or less what you might expect. If the docstring doesn't make it clear, it would be good to know.

edit: Also makes docs build on the latest Julia version.

edit-2: Also changes versions on which this is tested to 1.0 and 1, and disallows 1.3 failure on nightly. I'm guessing this just hadn't been updated for some reason.

[willtebbutt]

These tests cover everything covered in the to_vec tests other than Dict, which doesn't currently play nicely / I'm not entirely sure how best to construct a tangent, nor how to add it to the primal.

[willtebbutt]

CI failure is weird, pretty sure it's a one-off.

edit: by which I mean it's ready for review

[willtebbutt]

Looks like the Manifest probably just needed re-generating -- it uses a relative path to get the correct version of FiniteDifferences. Does anyone know if this is the right way to do things (still?)?

[oxinabox]

It is afaik.

[nickrobinson251]

one small suggestion. otherwise LGTM - thanks!

[willtebbutt]

@nickrobinson251 this PR no longer exports difference or rand_tangent. Are you on board with this? I'm keen not to document / export this functionality for now because we don't strictly need it until #97 is closer to completion. It's possible that we'll realise something here isn't quite right, and I would prefer not to force us to make multiple minor releases.