DmdAnal: Speed up the first demand analysis run by not recording used-once information
I don't really see how we use used-once/single-entry information (e.g. the 1* part of demand signatures) after the first round of demand analysis. Well, maybe a bit for eta-expansion of thunks, not sure.
The used-once information is quite inefficient to compute compared to a plain strictness+absence analysis. I think we could parameterise demand analysis by analysis mode, which would allow us to avoid inefficient postProcess* calls and whatnot in all but the very last demand analysis run, where we actually exploit used-once information for identifying single-entry thunks.
If we see that we actually lose performance, we can still do the full analysis in -O2 and plain strictness+absence in -O1.