Allow avoiding recompilation

[rikhuijzer]

This PR adds a keyword argument use_distributed which can be disabled to disable parallel processing. Instead, everything will be executed in the same process which avoids recompiling for each process. This can be more quick in systems with few threads such as GitHub Runners depending on the workload.

Also, rename run_notebook to run_notebook! because it mutates the first argument.

[rikhuijzer]

0a21852 manages to avoid recompilation completely. All tutorials are evaluated in the same process:

Sampling 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| Time: 0:00:20
106.714471 seconds (237.59 M allocations: 22.918 GiB, 4.17% gc time, 1.20% compilation time)
[ Info: for linear_regression.jl
[ Info: Starting evaluation of Pluto notebook logistic_regression.jl
  Activating environment at `~/git/TuringGLM.jl/docs/Project.toml`
   Resolving package versions...
  No Changes to `~/git/TuringGLM.jl/docs/Project.toml`
  No Changes to `~/git/TuringGLM.jl/docs/Manifest.toml`
┌ Info: Found initial step size
└   ϵ = 0.003125
Sampling 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| Time: 0:00:27
 41.427996 seconds (42.97 M allocations: 3.075 GiB, 1.58% gc time, 10.62% compilation time)
[ Info: for logistic_regression.jl

notice only 3 GiB allocations on the second tutorial

[rikhuijzer]

I've tested this on TuringGLM's tutorial building. Everything appears to be fine