QU demodulated timestreams should have dedicated pointing and projection optimizations

[mhasself]

The case of demodulated timestreams is special because the Q and U pseudo-detectors share pointing precisely. Updating the signal or weights matrices for both pseudo-dets will save a pointing computation and optimal for cache. (Detector pairs could also be treated this way, if you're so lucky to always have them.)

For that matter it may be worth exploring binning signal and weights maps at the same time, to maximally reuse the pointing computation.