Speed up multiple-class gradient calculations #162
Description
Description of feature
It should be possible to speed up gathering gradients for multiple classes of the same sequence by re-using the one gradient pass of predicting all classes, rather than doing it separately for every class. We had a preliminary attempt at this in #84, but it was demerged in #104 because of issues.
It's been the last remaining thing open in #114 which otherwise addresses long-fixed issues, so I'm splitting it out into a new issue. Since the original PRs, I've reworked/cleaned up the explainer backend, so it's probably worth going back and trying to see if we can get this to work again.