feat: improve sample efficiency

[lsorber]

Changes:

1. Conformal Tights now fits the wrapped estimator on all available data, to be used for point prediction. As a result, point predictions will generally be more accurate than before.
2. A base estimator is introduced (cloned from the wrapped estimator) and trained on the training set (which is a subset of all available data), to be used as the center for conformal prediction. A consequence of (1) is that the model's point predictions may themselves have low probability according to the calibrated conformal predictions.

Updated README images: