Optimize ScalarQuantizer

[Ray0907]

RS_quantiles with nth_element

Replace O(n log n) std::sort with O(n) std::nth_element for quantile
computation. This addresses the TODO comment requesting quickselect.

Changes:

• Use std::nth_element to find lower and upper quantile positions
• Eliminates unnecessary full sorting of x_copy vector
• Maintains identical behavior for quantile calculation
• Theoretical performance O(n log n) → O(n)

[meta-codesync]

@limqiying has imported this pull request. If you are a Meta employee, you can view this in D86234649.

[subhadeepkaran]

feels like a corner case that needs to be handled here when o > n/2
then it opens the scenario for being vmin > vmax

i think we should ensure: o = std::clamp(int(rs_arg * n), 0, n/2);

i can put out a follow-up diff on this:

[Ray0907]

You're right there's an issue. However, the correct bound should be (n-1)/2 instead of n/2 to ensure vmin_index <= vmax_index (i.e.,
o <= n-1-o). I've updated the fix accordingly.