Working reverse AD and more efficient forward-backward

[gdalle]

Writing down the computations allowed for faster code and working chain rules

• Add one page of docs on scaled forward-backward and its gradients
• Add working loglikelihood gradient computation in ChainRulesCore extension
• (breaking) Move alpha and beta out of public API
• Add more tests of interface and correctness
• Stop testing on nightly

[codecov]

Codecov Report

Patch coverage: 98.19% and project coverage change: +0.30% 🎉

Comparison is base (bca1bab) 93.16% compared to head (dd8714c) 93.47%.
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #33      +/-   ##
==========================================
+ Coverage   93.16%   93.47%   +0.30%     
==========================================
  Files          19       18       -1     
  Lines         483      521      +38     
==========================================
+ Hits          450      487      +37     
- Misses         33       34       +1     

Files Changed	Coverage Δ
src/inference/viterbi.jl	95.45% <ø> (ø)
src/types/abstract_hmm.jl	90.90% <ø> (ø)
src/inference/forward.jl	92.15% <88.23%> (ø)
ext/HiddenMarkovModelsChainRulesCoreExt.jl	100.00% <100.00%> (ø)
src/HiddenMarkovModels.jl	100.00% <100.00%> (ø)
src/inference/baum_welch.jl	90.69% <100.00%> (ø)
src/inference/forward_backward.jl	96.93% <100.00%> (+1.35%)	⬆️
src/types/abstract_mc.jl	97.29% <100.00%> (+0.07%)	⬆️

... and 1 file with indirect coverage changes

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.