Preserve correlation for nonconstant small-angle samples

[antonsoo]

Valid small-angle samples are rejected as degenerate because the Fisher–Lee and Jammalamadaka–SenGupta implementations compare positive variance terms with an absolute np.isclose(..., 0) tolerance. For example:

import numpy as np
from pycircstat2.correlation import circ_corrcc

a = 1e-5 * np.array([0., 1., 4., 2., 3.])
print(circ_corrcc(a, a, method="fl"))
print(circ_corrcc(a, a, method="js", test=True))

Both currently raise despite nonconstant input and unit self-correlation. Accept positive representable denominator and fourth-moment values. Keep constant and antipodal inputs degenerate by checking raw-angle sine residuals against argument-scale rounding, rather than imposing a fixed variance floor.

The new regressions cover small nonconstant samples, both inference modes, nonzero constants, antipodal directions and full-turn aliases. Nine configurations fail on unchanged main; all 28 pass afterward. The full suite passes 394 tests with hea==0.1.2; the fatal Flake8 gate passes. The existing latest-Hea collection limitation noted in #23 is unchanged.

A separate workload uses unmodified recorded KITTI OXTS roll/pitch/yaw windows. All 168 public calls that previously failed now agree with independent scalar pair sums/rescaled-moment references within 7.4e-11. Nine ordinary bundled-data controls remain exact. These checks establish numerical evaluation, not calibration of inference on serial sensor data. #21's current head still reproduces the failures. This is separate from #23's jackknife standard-error correction; extremely small products can still underflow.

[huangziwei]

will go through your patches next week. I'd like to upgrade hea with polars 2.0 first and cleared all upstream compatibility first. thanks for the help.