Interleaving pos/neg prompts to match what the repe code expects

[chanind]

It appears that the original Repe reading vector code expects every positive example to be paired with its corresponding negative example next to it in a big list passed to the find_directions() method. This isn't documented anywhere, but means we can't simply pass prompts to the find_directions() method unimpeded. Also, the Repe PCA code doesn't properly reflect each difference vector across the origin, so it's important to have a balance of <neg, pos> and <pos, neg> pairs passed to the find_directions() method.

This PR addresses these issues by adding a multi_answer_method param to the RepeReadingControl algorithm. By default, this will select just the first incorrect answer to pair with the correct answer, but can also be set to random_incorrect to pick at random, or repeat_correct to instead duplicate the correct answer to pair with every incorrect answer. This PR also alternates the order of the <neg, pos> for each example to try to ensure a good balance of directions for PCA.

[chanind]

Merging this now for the sake of speed to test it on colab. @dtch1997 Please feel free to code-review this still, any changes can be addressed in a follow-up PR