To obtain constant GM the bias current Id or Ic for transistors should be proportional to absolute temperature, Iptat.
Exactly, and doing exactly that was an early innovation in bipolar circuit design.
Why isn't the same thing done in their example? Is there something about CMOS which makes it untenable?
Thinking about it now, I never noticed such a thing in published CMOS operational amplifier designs but I assumed like later published bipolar designs, they just left that part out.