Performance: rows/finalizabledata: do finalize and transpose in one pass

[chriseth]

In the block and VM machine we have rows of cells that are writable. At some point their range constraints are deleted and only the value is stored (together with a flag to denote that the value is known). This process is called "finalization". At the end of witgen, the values are then stored into column vectors (while we had rows before), which is called transposition.

There is probably no real need for storing finalized rows and we could just do transposition at the point of finalization. We could do it less often than finalization to reduce the impact of memory cache misses.