Enable parallelization

[sethaxen]

This PR sets up internal usage of Transducers. The main benefits of this are to allow fine-grained control of parallel execution. With this PR both ELBO computation and multi-path Pathfinder are run with multithreading by default (when compatible with the user-provided RNG), but this can be customized by the user to be distributed, sequential, etc.\

This PR supersedes #11.

[codecov]

Codecov Report

Merging #38 (518c3c4) into main (972e0ef) will increase coverage by 0.05%.
The diff coverage is 100.00%.

❗ Current head 518c3c4 differs from pull request most recent head 27e542a. Consider uploading reports for the commit 27e542a to get more accurate results

@@            Coverage Diff             @@
##             main      #38      +/-   ##
==========================================
+ Coverage   99.10%   99.15%   +0.05%     
==========================================
  Files          11       12       +1     
  Lines         335      357      +22     
==========================================
+ Hits          332      354      +22     
  Misses          3        3              

Impacted Files	Coverage Δ
src/Pathfinder.jl	100.00% <ø> (ø)
src/elbo.jl	100.00% <100.00%> (ø)
src/multipath.jl	96.66% <100.00%> (+0.51%)	⬆️
src/mvnormal.jl	100.00% <100.00%> (ø)
src/optimize.jl	100.00% <100.00%> (ø)
src/singlepath.jl	100.00% <100.00%> (ø)
src/transducers.jl	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 972e0ef...27e542a. Read the comment docs.

[sethaxen]

Until JuliaFolds/Transducers.jl#521 and JuliaFolds/Transducers.jl#522 are resolved, we would need to accept type piracy and instability, so for the moment, we'll just use Transducers to parallelize multi-path Pathfinder (and leave ELBO estimation un-parallelized)

By introducing a mutating pattern in a few key places, type-stability is rescued.