CIs based on tail bounds

[nhejazi]

We should implement several alternative methods for inference, not drawn directly from the normal approximation to the asymptotic distribution of influence functions. At least two classes of options are available

1. Use of a logistic distribution (or other heavy-tailed distribution) as reference instead of normal/t (post-hoc)
2. Bernstein, Bennett, or Hoeffding bounds for influence function-based estimators, as per https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2827893/