I suspect the apparent effects of these different training programs are mostly down to selection rather than any specific benefit of the training programs themselves.
Entry to MSTP is limited to people who have shown a strong interest in and aptitude for science from an early age. The same is likely not true for the overall pool of MD-only and PhD-only grads.
F grants and K grants are also competitive, and require the applicant to write a very specific and detailed proposal that will be examined with a fine-tooth comb by a review committee of unaffiliated established scientists.
The T32 at my institution was pretty much not competitive at all. All I had to do to apply was hand in a two-page research proposal, and the relevant faculty acted like admission was pretty much a foregone conclusion. The other people in the T32 program while I was there ran the gamut from very focused and dedicated PhD-only postdocs to a couple of MD-onlys with a vague interest in academics who seemed mostly to be testing the waters.
Thus it would make sense that the pool of people who receive MSTP, F series, and K series funding would be more successful than the pool of people receiving T32 funding, regardless of the actual training received in each of those programs. As you can see the bulk of the people who finish T32 programs never submit an application for an RPG, vs people with Fs and Ks who mostly do.
There's not a good way to control for this, although if you could find a pool of people who were admitted to MSTP but chose to go to a straight-PhD program, that could be illuminating. That would be a tiny tiny group though, and I think stochastic effects (which are very large at this level of competitiveness) might overwhelm the between-group effects at that point.
Edited to say: I'm not saying that training doesn't matter, I'm saying that getting good or bad training is more specific to the environment, the mentor, and the project than to the funding mechanism for the trainee.