more stable yeo johnson roundtrips

[mathause]

• Addresses the issue mentioned in reformulate lambda function #476 (comment)
• Tests added
• Fully documented, including CHANGELOG.rst

With this we can take another stab at #476. It also seems a bit faster. Modeled after scikit-learn/scikit-learn#26188

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 50.11%. Comparing base (9b0b76b) to head (3c12f01).
Report is 127 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff             @@
##             main     #494       +/-   ##
===========================================
- Coverage   87.90%   50.11%   -37.80%     
===========================================
  Files          40       50       +10     
  Lines        1745     3512     +1767     
===========================================
+ Hits         1534     1760      +226     
- Misses        211     1752     +1541     

Flag	Coverage Δ
unittests	50.11% <100.00%> (-37.80%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

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

[veni-vidi-vici-dormivi]

Huh interesting, so that means that np.power yields less precision than np.exp and np.log, right? That's good to know.

[mathause]

Thanks. I think its more the np.expm1 and np.log1p. From the docstrings:

• expm1: This function provides greater precision than exp(x) - 1 for small values of x.
• log1p: For real-valued input, log1p is accurate also for x so small that 1 + x == 1 in floating-point accuracy.

(interesting the names are not symmetric...)

[veni-vidi-vici-dormivi]

Interesting. Thanks for doing that!