fix(covmodel): ensure positive percentile_scale using bounded brentq

[n0228a]

fix(covmodel): ensure positive percentile_scale using bounded brentq

Problem

CovModel.percentile_scale(per) returned negative distance scales (e.g. -6.215 for SuperSpherical with nu >= 2). A percentile scale represents physical distance and must be positive ($r \ge 0$).

Cause

• Correlation functions are even $\rho(-r) = \rho(r)$, making $1 - \rho(r) - \text{per} = 0$ symmetric with roots at $+r_0$ and $-r_0$.
• percentile_scale used unconstrained scipy.optimize.root.
• For models with sharp drop-offs (SuperSpherical with $\nu \ge 2$), flat gradients near $x_0 = 0.9 \cdot \ell$ caused the solver to overshoot across zero and converge onto the mirror root $-r_0$.

Fix

Replace scipy.optimize.root with scipy.optimize.root_scalar(method="brentq") bounded on $[0, b]$:

• Constrains the search to $r \ge 0$.
• Dynamically expands upper bound $b$ until $\text{curve}(b) > 0$.
• Guarantees convergence without derivative sensitivity.

Verification

import gstools as gs
for nu in (0.5, 1.0, 2.0, 5.0):
    m = gs.SuperSpherical(dim=2, nu=nu, len_scale=10.0)
    print(nu, m.percentile_scale(0.9))

• Before: nu=2.0 -> -6.215, nu=5.0 -> -4.575
• After: nu=2.0 -> +6.215, nu=5.0 -> +4.575

Test suite: pytest tests/test_covmodel.py (7/7 passed).