I see...well, i think there have to be best practices to do this. I just don't know them. Fact is, that the default collision detection methods of jPCT-AE alone won't be sufficient. I have two ideas how to deal with this and i'm not sure, if they are applicable...however, my ideas usually aren't thaaaat bad, so maybe they'll work...
Idea one: Don't use real collision detection at all, but a kind of heuristic based on player's positions and actions. Like: If player 1 faces towards player 2 and player 2 doesn't block and player 1 does a kick and the distance of both is < x, then the kick of player 2 will hit player 1. This can be checked regardless of any animation playing and it's very fast to do. It can be tricky to tweak it to look reasonable though.
Idea two: Don't do collision detection on the models but on some abstract and simplified model. Like an ellipsoid for the players and some spheres that marks the end of each arm and leg. If a player kicks or punches, move the spheres so that their positions reflect the animation playing (more or less) and check collisions between the spheres and the player's ellipsoid...something like that. That's harder to do than the first approach IMHO. Especially you have acutally two "animations" to play with one being the animation of the actual view and the other one the "animation" inside the collision model. But it *might* produce better results.
In both cases, the md2 models and animations only represent the view, not the data on which the actual collision detection is based on. You won't find a magic method inside jPCT-AE that handles this. It's a special case related to fighting games and you have to solve this on your own. However, maybe one of these ideas is helpful.