improve numerical stability of vector.rb

[alshedivat]

when qrot[r, r] is close to 0, ruby might represent it with a small negative value (e.g., -2.8475347e-16), which results in error when passed to Math.sqrt. adding a small epsilon constant to qrot[r, r] inside Math.sqrt solves the issue.

[cardmagic]

FYI, the original classifier gem (from which classifier-reborn was forked) already addresses this numerical stability issue with a slightly different approach:

singular_values = q_rotation_matrix.row_size.times.map do |r|
  Math.sqrt([q_rotation_matrix[r, r].to_f, 0.0].max)
end

Instead of adding an epsilon, it clamps negative values to zero using [value, 0.0].max before taking the square root. This avoids artificially inflating zero singular values while still preventing the Math.sqrt error from tiny negative floating point values.

See: https://github.com/cardmagic/classifier/blob/master/lib/classifier/extensions/vector.rb#L110-L112