Not sure if this is best, but it fixes the issue of upsampling lower orders to
higher orders multiple times (mixing lower order to higher order, cutting back
to lower order by dropping the extra channels, then upsampling again, applying
the HF scale multiple times and messing up the high frequencies/energy vector),
as well as having too-strong high frequencies of upsampled content in general.
As it is, with the new upsampling method, the lower order signal contributes to
some higher order channels, which helps offset the high frequency scaling
difference (except first- to second-order, which adds nothing).
This should allow for clearer and less diffuse responses. While a cube is in
the blindspots for second-order, resulting in an identify transform for first-
to-second-order, they do start contributing to third-order channels, which
should make for sharper virtual points. The fixed HF scales should also play
nicer with larger order upmixes.
