Playing Devil's Advocate...
The statement that frequencies above 20kHz don't matter rests upon the assumption that the ear is linear. If the ear is not linear (I don't know whether it is not not) then frequencies above 20kHz will matter, as the ear will be able to mix higher frequencies down to less than 20kHz. For example, if we have frequencies of 56kHz and 59kHz, the ear MIGHT be able to discern a difference frequency of 3kHz. No doubt this effect could be reproduced by signal with a sampling rate of 44.1KHz, but only if the analogue systems, before the sampling stage, reproduce any non-linearity in the human ear.
Incidentally, you can get speakers that create a localised beam of sound, that the person sitting next to you cannot hear. They work by transmitting frequencies above the audible range. These high frequencies can be beamformed by a relaitively small speaker array, so the sound is localised. They then rely on the non-linearity of the ear (or maybe the air around the ear?) to mix the ultrasonic frequencies down to audible frequencies. I guess there must be non-linearity in the human auditory system!
On the subject of 24-bits my understanding is that 16-bits is adequate, provided the levels (scaling) are set correctly in the recording. What 24-bits delivers is the ability to do a crappy job of the mixing, and still end up with the full dynamic range of the human ear. 24-bits is probably a temporary solution though, as manufacturers will engage in the usual Loudness War [1], and push the signal to the top of the dynamic range. Before long 24-bit audio will be equivalent to 16-bits (since the 8 least significant bits will be unused) and the next big thing will be 32-bit audio.
Having said all that, I'd guess that the speakers will be the limiting factor in most sound systems, not the recording format.
[1] http://en.wikipedia.org/wiki/Loudness_war