Make epi_cor() / summarizer compare apples to apples, and/or warn against usage

[brookslogan]

Users might reach for epi_cor() for feature selection. This might be okay to judge whether something seems anywhere plausible for use at all, but for judging exact/relative usefulness it's not the best...

• It doesn't provide a one-number summary of usefulness per signal-lag. (Also applies to even the anywhere-plausible judgment.)
• It doesn't compare apples to apples across signal-lags with different availability.
• It is not tailored to the evaluation metric of whatever model is actually going to be fit.
• Selection and regularization should probably be handled by a modeling package, not by ad-hoc code atop an EDA utility.

We should address the points above and/or add appropriate cautions in docs.