Add `MvNormal(::FiniteGP)` in a `@requires`

> It's not necessary to compute it if you don't call `rand` or `logpdf` shrug
>
> One approach that doesn't touch the structure of `FiniteGP` would be to add a method `MvNormal(::FiniteGP)` that returns the corresponding `MvNormal` distribution which could then be used instead of `FiniteGP` if you don't want to recompute the cholesky decomposition, e.g., if you call `rand` or `logpdf` multiple times.

Ooh, I like that. Maybe we have a `@require` in MeasureTheory that adds the `MvNormal` method. Then we can use the faster representation too, should work out great!

_Originally posted by @cscherrer in https://github.com/JuliaGaussianProcesses/AbstractGPs.jl/issues/232#issuecomment-947633806_

What do you think @mschauer?