Support for additional `celerite` kernels

[vandalt]

Hello,

As discussed in #353, I implemented the Matern32Term and SHOTerm from celerite within the RadVel gp framework. This PR contains:

• Tests for celerite kernels (comparing with the outputs of a celerite object)
• An abtract class for celerite kernels that has some of the code common to all celerite kernels
• Methods to calculate the covariance matrix of celerite kernel (mainly useful for testing, but also uniformizes the GP interface between celerite and other kernels)
• New kernel classes for Matern 3/2 and SHO terms
• CeleriteLiklihood.predict() is now more general to support other celerite kernels

I saw that there is a gp-multiply-kernels branch, so I was wondering if I should rebase against this one instead of master ?

Thank you!

[vandalt]

Converted to draft because I just realized there is a bug when Q < 0.5 because the real terms are array. I'll try to use the (unused) 1and 3indices of CeleriteKernel.real to work around that.

[vandalt]

The issue with Q < 0.5 is now fixed, I added a test covering this as well.

[vandalt]

Marked as draft to rebase on #359, which refactors the GP API.

[bjfultn]

Is this still a WIP?

[sblunt]

Yeah, we'll need to reimplement the celerite functionality based on the tinygp refactor in #363

[vandalt]

Unfortunately could not work on this lately. I'm hoping to have time this winter, but in the meantime anyone feel free to take over!