Fix von Mises-Fisher sampler

[devmotion]

Fixes #1423.

As already discussed in #1423, in the special case mu = [1, 0, ..., 0] no rotation of the "initial" sample is needed. I think skipping the rotation completely is preferable over a numerical workaround such as #1918.

I used the opportunity to also add comments and references to the existing code of the von Mises-Fisher sampler.

Edit: I switched to using LinearAlgebra.reflector! and LinearAlgebra.applyReflector!.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 86.26%. Comparing base (bf4f03e) to head (022da79).

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1930   +/-   ##
=======================================
  Coverage   86.26%   86.26%           
=======================================
  Files         146      146           
  Lines        8765     8769    +4     
=======================================
+ Hits         7561     7565    +4     
  Misses       1204     1204           

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[williamjsdavis]

Nice work! This is an elegant solution, and I agree that it's preferable to the fix I proposed in #1918. I'm happy to delete that PR if/when this PR is accepted.

[andreasnoack]

Can't we use the reflector machinery from LinearAlgebra instead of reimplementing it here. I.e. reflector! and reflectorApply!. I do realize that I did a poor job documenting the two, but the implementations should be robust and I think we avoid the rotate field.