Gen Tensor/NonTensor Mixed

[jeremylt]

Follow-up on #1735

The current parallelization strategy for non-tensor and tensor bases means that we cannot currently mix them in gen backends.

The fix isn't too bad - we need to make a version of the tensor operator that assumes t_id_y == 1 by decomposing t_id_x = a + b * P_1D. Same tensor contractions in 2D, but just different mapping to threads. For 3D we'll need a new template that extends the 2D approach in the natural way instead of using 2D slabs.

Its straightforward, but I wanted to do this separately so the PR for #1735 doesn't get too big.