PERF Speed up KNeighborsClassifier fitting - #34564

LGTM

LGTM, thanks!

The one thing I wanted to check on this change is that fmax/fmin and the ternary form are not equivalent in the presence of NaN, so I looked at whether that can matter here.

C99 fmax returns the non-NaN operand, whereas the ternary propagates:

np.fmax(1.0, nan)              # 1.0        -> NaN ignored
1.0 if 1.0 > nan else nan      # nan        -> NaN propagates

So if a NaN could reach find_node_split_dim, max_val could become NaN with the new code where it previously stayed finite, and the resulting spread would poison the split-dimension choice.

It cannot, as far as I can tell. NaN is rejected at every entry point into the trees:

BallTree(X_with_nan)                                  -> ValueError: Input contains NaN.
KDTree(X_with_nan)                                    -> ValueError: Input contains NaN.
KNeighborsClassifier(algorithm="ball_tree").fit(...)  -> ValueError: Input X contains NaN.

Infinities behave the same under both forms, so NaN was the only divergence. The change looks safe to me on that front.

Two notes:

• I could not reproduce the timings, since this needs a Cython build I do not have set up, so I am only commenting on correctness rather than confirming the speedup.
• The diff also adds an attribution line to 34187.efficiency.rst, which belongs to a different PR. Harmless, but it is unrelated to this change and might be cleaner in its own commit or left out.