5 wrap models predict interval

[pat-alt]

Addressed type piracy issue. Now using a separate method predict_region to compute conformal prediction regions, while the generic predict computes point predictions as always. It probably makes sense to make that distinction explicit, because:

• I'm not sure how/if set-valued/interval predictions fit easily into the standard MLJ paradigm (probably not).
• Even if there was a way to make that work, I'm no longer sure that's even desirable. For any model that has been calibrated for conformal prediction, you can always still also produce point predictions. So it seems reasonable to preserve that functionality and treat conformal prediction as a separate functionality.