Support cases where NxM > memory

[ytoren]

For the supported built in metrics (where matrix multiplication is possible) we support breaking down the caclulation to [row or block 1xN or nxM] %*% [full NxM matrix] similarity caclulations and stack the similarity matrix row-wise. This works only if the entire feature matrix (NxM) to fit in memory.

For cases where $NxM$ > memory we want to allow the matrix multiplication work "block against block". This means we will stack the similarity matrix both row wise (as we do today) and column wise.

• Figure out the math (as long as we don't have to break $M$ this should still work)
• Introduce a second range parameter (or a row-range couplet) which defaults to (1, nrow(X)) (which is the current calculation)
• Figure out the vertical stacking (block matrix construction?)