"For those of you that have GeForce cards, how useful have you found hardware T&L to be?"
Extrodinarily, though I work with 3D visualization and play games, the former benefits much more then the latter at this time.
"Nvidia claims that it takes a substantial amount of processing off the CPU freeing it to do other things (and hence increasing framerate), but, some cards that don't have hardware T&L like the Voodoo4/5 seem to perform comprably."
Most games don't support hardware T&L, which is needed for it to be used. The V4/5 don't perform comparably in games that utilize hardware T&L properly. Under MDK2 you can pick to run the same level of quality graphics for both hardware T&L boards and software(hardware T&L as of right now only supports the most computationaly intensive type of lighting, specular) based ones and when tested with specular lighting enabled for both the GF2 is roughly 300% faster, give or take.
"Also, the GeForce scales upwards a lot more on CPU's than it should, implying that that the hardware T&L isn't kicking in as much as it should be (if it were, framerates would be the same on a P2-300 as they would be on an Athlon 1Ghz)."
Most games devote about 25%-35% of CPU time to T&L, the rest is running the other portions of game code. Now think of it this way, if you have a 500MHZ processor and would push 100FPS with software T&L, hardware T&L will give you roughly 25 extra FPS, but a 1GHZ processor would likely gain an additional 50FPS because it can process the game code twice as fast as it still has no need to handle T&L. This makes it appear as if scaling happens moreso on hardware boards then software which can be argued, but it is because T&L is in fact working so well. If this didn't happen then T&L would likely be a bottleneck.
The real benefits will come when developers use the extra power that T&L can offer. MDK2 starts to scratch the surface, but it still only utilizes additional lighting, not added poly counts which overall would probably make the game look even better. Evolva is another title that I would suggest you check out if you want to see what some of the new features on that GF2 MX of yours has, it utilizes not only additional polygon complexity but also Dot3 bump mapping. May not be everyone's thing in gameplay(it is odd, though I like it), but the grapihcs with either a GF based board or Radeon are very impressive.