Faster octree nearestKSearch

[mvieth]

Inspired by the logic in organized.h: always keep point_candidates sorted, inserting new entries at the correct places. This avoids repeated calls to std::sort and having more than k entries in point_candidates.

Benchmarks: NormalEstimation (without OpenMP) with octree:	NormalEstimation	Old	New
k=50, cloud=milk	2607 ms	1842 ms
k=50, cloud=mug	2043 ms	1505 ms
k=100, cloud=milk	5650 ms	3075 ms
k=100, cloud=mug	3824 ms	2436 ms

Calling nearestKSearch on every (valid) point of a cloud:	Search	Old	New
k=1, cloud1	818 ms	509 ms
k=1, cloud2	791 ms	613 ms
k=1, cloud3	773 ms	744 ms
k=5, cloud1	914 ms	637 ms
k=5, cloud2	950 ms	780 ms
k=5, cloud3	1083 ms	1033 ms

[larshg]

Nice improvements!