refactor tests into multiple Github CI steps

[yebai]

Similiar to TuringLang/JuliaBUGS.jl#149.

Let's remove the complex macros including numerical_testset, and turing_testset etc. Instead, let's simply group tests by files and run these test files in different CI runners.

Useful resource for local test: https://gist.github.com/torfjelde/62c1281d5fc486d3a404e5de6cf285d4

[yebai]

For HMC sampling correctness, we only need to test against one AD backend, but we currently test against a few.

We only need to test gradient evaluation for the AD functionality test, but we currently run HMC, which is an overkill.

[torfjelde]

Very much agree with this 👍

[torfjelde]

We only need to test gradient evaluation for the AD functionality test, but we currently run HMC, which is an overkill.

A counter point to this: TuringLang/DynamicPPL.jl#612

Sometimes AD only breaks when we're sampling 😬

[yebai]

Fixed by #2237