More efficient implementation of median

[gkronber]

The current implementation of median() creates an array and uses Array.Sort to find the median element. Sorting is however not stricly necessary as the median (or any quantile) can be determined using an O(n) algorithm.

[gkronber]

• In place implementation that alters the array: see §8.5 select(), "Numerical Recipes in C"
• Non-destructive selection (is 10x slower than select()): see §8.5 selip(), "Numerical Recipes in C"
• For k << N (e.g. the 10 smallest or largest value in an array of 1000 elements): Use Heapsort "make a single pass through an array of length N while saving the m largest elements." "Only log m, rather than m, comparisons are required every time a new element is added to the candidate list."

[gkronber]

r13033:

• implemented O(n) algorithm for median (and general quantiles) determination.
• Comparing output of new implementation with naive implementation (as well as performance) in the unit test
• let's see if all samples still produce the same results...

merge this after #2488

[mkommend]

r13059: corrected spelling mistake in median implementation.

Reviewed r13033.

[gkronber]

r13150: merged r13033 and r13059 from trunk to stable branch