[BUG] Brute-force kNN can return nonzero distances for identical vectors #2883

#2841 removed a rule that forced small distances to zero when two vectors had equal squared lengths. This addressed a real problem: distinct nearby vectors could incorrectly receive zero distance, affecting neighbor ordering and clustering.

As discussed in the PR, removing that rule also exposes underlying rounding errors when comparing a vector with itself. Negative errors are still clamped to zero, while positive errors remain. Taking the square root makes these errors more visible with the Euclidean metric. The PR documents this tradeoff and adjusts the Go brute-force test tolerance accordingly.

Observed behavior :
In a float32 test with 800 vectors of 32 dimensions, searching an independently allocated copy of the dataset returned nonzero self-distances for 256 vectors:

• Maximum squared Euclidean self-distance: approximately 0.000732.
• Maximum Euclidean self-distance: approximately 0.0271.

This occurred with both k=8 and k=65. The correct self-neighbor IDs were returned. Similar behavior occurs with dataset slices, column-major inputs, and filtered searches.

Expected behavior :
Identical finite vectors should have zero distance, including when queries are copies or slices of the dataset.
Could we preserve this behavior while retaining the improvements from #2841? This would complement the diagonal correction already added for certain pairwise-distance calls, extending support to brute-force kNN.