Improve Handling of Fill-Values in RAD_MAX Estimator

[RuneDominik]

In the current version, the RAD_MAX estimator does not handle fill-values and thus produces results that inter-/extrapolate between actual values and fill-values, yielding results that do not resemble the expected behavior in edge-cases (e.g. RAD_MAX values sharply rising after some energy, see Fig. below). This PR introduces some form of combinatoric edge-case handling in 1 and 2D grids.

rad_max_cuts_new.pdf

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 95.74%. Comparing base (cb92eef) to head (b6b33c9).
Report is 2 commits behind head on main.

❗ Current head b6b33c9 differs from pull request most recent head 26a9d8b. Consider uploading reports for the commit 26a9d8b to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #282      +/-   ##
==========================================
+ Coverage   95.60%   95.74%   +0.13%     
==========================================
  Files          62       62              
  Lines        3278     3381     +103     
==========================================
+ Hits         3134     3237     +103     
  Misses        144      144              

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

[maxnoe]

You are checking for BaseNearestNeighborSearcher here twice, only the check on target_simplex is added, but then you do the same in both cases.

So just checking for issubclass(BaseNearestNeighborSearcher should be enough?

[RuneDominik]

target_simplex == -1 means target_point outside grid, thus this checks for cases, where extrapolation will actually be applied and only in these cases the issubclass check on self.extrapolator.__class__ yields the needed information. In the second case, thus target_simplex != -1, this needs to check self.interpolator.__class__.

Basically, this accounts for the possibility that the user mixes and uses NearestNeighbor for extrapolation and actual interpolation otherwise (and vice-versa) at the same time.

[maxnoe]

Why can't we just check for nearest neighbor? Why this complex setup?

[RuneDominik]

That setup is to catch cases where the user decided to mix approaches and use e.g. NearestNeighbor for interpolation but an actual extrapolator for extrapolation or vice-versa. I'll extend the comment in code to make this more clear for the future.