Fix nondeterministic quadrature position bug

While originally intended to silence #110 (closed), it turned out that there was nasty little bug in the code that preevaluates quadrature data in the constructor. There were two statements writing to the same temporary using quadrature formulae of different size. If by coincidence the correct one was last, we got correct results, otherwise they would be slightly off (though not in the range that makes a test fail). The patch fixes this by always picking the correct quadrature formula.

This fixes #110 (closed) .