f_clash optimization for speed

[nejci]

1. Use squared Euclidean distance Instead of full Euclidean distance between atoms - avoid computing sqrt. In function getRclash v kernels/classification.cl return squared distances (e.g., 2.5 -> 6.25 etc.). Add new function lengthNormSq3 in kernels\vector3.cl, that returns squared Euclidean distance.
2. Optimize memory access by transposing atom data to speed up scanning through atom pairs.
3. Efficient computation of sum of squared distances: https://www.geeksforgeeks.org/sum-of-squares-of-distances-between-all-pairs-from-given-points/

[nejci]

Great @te2445! Now we have points 1.) and 2.) covered 🎉