fix: IntervalEncoder average-method window predicate (OR should be AND)

[AlejandroCoronadoN]

IntervalEncoder(method="average") never applies the lower bound of its averaging window. In _mk_average the predicate is:

predicate = (xs < (interval + span)) | (xs < (interval - span))

Since interval - span < interval + span always, the second condition is a subset of the first, so the OR collapses to just xs < (interval + span): there is no lower bound. Every point below the interval (however far) leaks into the weighted average, which biases the smoothed height low.

The docstring describes span as "the span around the interval" (a symmetric, two-sided window), and the sibling method="normal" uses all points weighted by a Gaussian, so average is meant to be the hard-windowed version of the same idea. The fix uses AND:

predicate = (xs < (interval + span)) & (xs > (interval - span))

Repro with a linear target y = 2x (the smoothed height at each quantile should track 2 * quantile):

import numpy as np
from sklego.preprocessing import IntervalEncoder

np.random.seed(0)
x = np.random.uniform(0, 10, 2000)
ie = IntervalEncoder(n_chunks=5, span=0.1, method="average").fit(x.reshape(-1, 1), 2 * x)
print(ie.heights_)   # before: biased low (~17.5 at the top quantile near 10); after: ~20

Added a regression test. The existing method="average" test only used a constant target, which cannot detect the bias (averaging any subset of a constant array still gives that constant).

[FBruzzesi]

Thanks @AlejandroCoronadoN that indeed seems like a copy-pasting gone wrong + wrong logical operator! Thanks for spotting and fixing it