Rework orthogonalization in Lanczos without direct Torch access

[Cdebus]

The dndarray __getitem__ function appeared to be rather slow in the Gram-Schmitt orthogonalization of the Lanczos algorithm (linalg.py):

for i in Lanczos_Iterations:
          ...
          for j in range(i)
               vr = vr - projection(vr, V[ : , j ])

--> ca 60s for 300 iterations for __getitem__ Using only the item access in each of the 300*(300+1)/2 iterations yielded a computation time of roughly 40s

for i in Lanczos_Iterations:
          ...
          for j in range(i)
               temp = V[ : , j ]

Also the projection(a,b) function used item access and was rather slow.

projection(a,b) = (dot(a, b) / dot(b, b)) * b

To speed this up, an arithmetic solution was implemented, that calculates parts of the dot-products on the the local torch level and uses allreduce to gather them. However, this is not nice at it reveals low-level communication to the user. We should think of a way to abstract this. @Markus-Goetz @coquelin77