Correct the Fisher-Lee jackknife standard error

[antonsoo]

The Fisher–Lee jackknife test uses the variability of the leave-one-out estimates without converting it to the variability of the pseudovalues. This makes its confidence interval too narrow by a factor of n - 1.

A reproducible example uses the bundled Milwaukee wind measurements:

import numpy as np
from pycircstat2 import load_data
from pycircstat2.correlation import circ_corrcc

data = load_data("milwaukee", source="jammalamadaka")
a = np.deg2rad(data["theta"].to_numpy())
b = np.deg2rad(data["psi"].to_numpy())
print(circ_corrcc(a, b, method="fl", test=True).reject_null)

This prints True before the change and False afterward. The correlation remains 0.19103419343618408. Independently forming the delete-one pseudovalues gives a normal 95% interval of approximately [-0.08164, 0.46528], which contains zero.

Use (n - 1) * sqrt(var(leave_one_out, ddof=1) / n) for the standard error. The bias-corrected center, normal critical value, coefficient and other correlation methods stay unchanged. This follows the pseudovalue variance calculation in Sawyer's jackknife notes, equations 1–3.

The new measured-data regression fails before the change. The full suite passes 367 tests with hea==0.1.2, and the workflow's fatal Flake8 checks pass. The latest hea==0.1.10 currently fails collection on both unchanged main and this branch because regression.py imports its removed lm export. Separate scalar pair-sum/pseudovalue calculations verify 136 measured-data and transformed-control cases; 80 incorrect decisions are corrected, while coefficients and other-method outputs remain exact. The same reproduction still fails on #21's current head.