[WIP, not to merge] Migrate to Enzyme

[simsurace]

The library is getting hard to maintain mainly because of the way it works around Zygote limitations, which is by using a lot of hacks and type piracy that are relying on various package internals and are therefore prone to breakage.
This is a draft to migrate to Enzyme and getting rid of all AD-specific hacks.
TODOs:

• Make examples run with Enzyme
• Add integration tests for AD to cover all input types and kernels.

[coveralls]

Pull Request Test Coverage Report for Build 7863797563

• -1 of 102 (99.02%) changed or added relevant lines in 11 files are covered.
• No unchanged relevant lines lost coverage.
• Overall first build on drop-zygote at 92.95%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/models/linear_gaussian_conditionals.jl	16	17	94.12%

Totals
Change from base Build 7625236479:	93.0%
Covered Lines:	1002
Relevant Lines:	1078

---

💛 - Coveralls

[simsurace]

The first example, exact time learning, now runs on Enzyme. There might be a bug still in Enzyme, which is why the shadows wrt. the inputs are needed right now.

[willtebbutt]

@simsurace are you still interested in this? I've recently excised basically all of the AD-related code in favour of just using Mooncake.jl, so I imagine you would just need to check that Enzyme successfully runs on some high-level examples?

[simsurace]

Hi, nice work on Mooncake and great that we have another option now. I'm still interested but since my work currently does not need GPs I have had to set these interests somewhat aside. I will almost certainly be using GPs again sooner or later so I hope JuliaGPs lives on.

[willtebbutt]

Thanks @simsurace. In this case I'm going to close this PR for the time being, as I suspect that things have moved on quite a bit since you were attempting to get it to work. Please feel free to reopen at some point in future if you start using GPs again!