Uses only the first 100 rows (not a random subsample) to keep the
O(n^2) pairwise-distance computation cheap, matching the upstream
implementation exactly. Multivariate input is measured by the squared
Euclidean distance between rows, as in the upstream get_kernel_width().