Optimize QCombiner

[mtfishman]

The QDense combine and uncombine functions are using a slow QDense constructor that determines the output tensor from the divergence here. One should be able to determine the non-zero output blocks from the non-zero input blocks, like in the new version of QDense contraction. Right now, for high ordered tensors with a lot blocks, those constructors are taking up a vast majority of the time of those functions.

In addition, they should be using the new uninitialized QDense constructor, which should also speed things up even more (once the QDense constructor is replaced).