Renaming some things

[willtebbutt]

Following @GiggleLiu's enthusiasm for renaming Composite to Tangent in this Zygote issue, I think it might be a good time to consider doing some renaming. The impression I get is that using Tangent in the names of things is probably a good idea, and could avoid some serious confusion.

@oxinabox discussed this briefly on slack, and he made the following suggestions:

• Composite -> CompositeTangent (i.e. a tangent that's built out of other stuff)
• DoesNotExist -> NoPossibleTangent
• Thunk -> ThunkedTangent
• Zero -> ZeroTangent

We might also want to consider whether AbstractDifferential is a sensible name, or whether we want to rename that to AbstractTangent or something.

@GiggleLiu do these proposals seem intuitive?

Then there's the issue that we're actually using these to represent cotangents whenever we're doing reverse-mode, which hasn't seemed to be problematic so far. Maybe @sethaxen could comment on whether we want to take another look at this if we're doing some renaming?

[GiggleLiu]

Thanks for taking this seriously.

@GiggleLiu do these proposals seem intuitive?

Definitely, all these renaming looks good. Since this is a big change that may affect packages depending on ChainRules, maybe it is better to make it slow, that is, showing a deprecation warning for several minor versions.

Add: at least the Composite deserves the renaming. When one says, "ChainRules returns an object of Composite type" as the gradient, many people might think the Composite type is a general purposed composite type rather than a specific one.

[willtebbutt]

Yup, we would be sure to deprecate sensibly -- as you say, it would almost certainly break a lot of code!

[oxinabox]

Since this is a big change that may affect packages depending on ChainRules, maybe it is better to make it slow, that is, showing a deprecation warning for several minor versions.

We will just make all the follow up PRs ourself.
As well as deprecating for 1 version.