maybe easier to go the granular route than mess about with fft. i just did a quick search for 'realtime timestretch' and a lot of people are saying that the granular based patches/plugins are good.
makes sense too....cos if you use a phase vocoder approach, you will have to speed thru the audio to get back from a delayed signal to a realtime one...but with granular, you can just merge grainclouds.