sklearn euclidean_distances: 9% error for close points far from 0; center first

scikit-learn 1.9.1, NumPy 2.5.4, float64. euclidean_distances (and pairwise_distances(metric='euclidean'), which calls it) uses ||x||^2 + ||y||^2 - 2 x.y, so the relative error is about eps*||x||^2/d^2 (max over pairs ~8x that estimate here). 10-D points with offset 1e4 and spread 1e-3: max relative error 9.2e-2 vs scipy cdist; offset 1e3: 1.0e-3; offset 1e6, spread 1: 9.9e-4; centered data: ~1e-15. Subtracting the column mean of X from both X and Y before the call brings it to 1e-15 at no cost, and scipy.spatial.distance.cdist (direct differences) is accurate without it. The same applies to the hand-written matmul trick. For float32 input sklearn upcasts in chunks and was accurate to float32 rounding (8.8e-8) on centered data, but casting uncentered data with offset 1e4 to float32 already destroys the 1e-3 spread before any distance is computed.