fix: use robust pseudo p-value for two-sided significance

[madhavcodez]

Title

fix: use robust pseudo p-value for two-sided significance

Summary

The two-sided branch of _permutation_significance in esda/significance.py
derived the p-value from percentiles of the reference distribution. When the
conditional reference distribution is constant (every permuted value is
identical), those percentiles collapse to a single value, so both the lower and
upper tail counts include the entire reference distribution. The resulting count
exceeds p_permutations, and the p-value can exceed one.

>>> import numpy as np
>>> from esda.significance import calculate_significance
>>> calculate_significance(5.0, np.full((1, 19), 5.0), alternative="two-sided")
1.95

This replaces the percentile approach with the equivalent robust pseudo
p-value suggested in the issue:

2 * (min(greater, lesser) + 1) / (permutations + 1)

clipped at one. The result stays in (0, 1] for degenerate nulls and matches
the existing one-sided counting conventions used by the greater/lesser
branches, so the directed <= two-sided invariant still holds.

Changes

• esda/significance.py: two-sided branch now uses the clipped pseudo p-value.
• esda/tests/test_significance.py: regression tests covering
  • the degenerate constant null with the statistic on the constant, asserting
    the result is exactly 1.0 for both the scalar and vector inputs;
  • the second failure mode of the old formula, a constant null with the
    statistic off the constant, asserting the pseudo p-value 0.1;
  • a seeded normal null, asserting the exact pseudo p-value 0.016;
  • the two-sided == 2 * directed identity on a one-sided statistic.

Testing

• pytest esda/tests/test_significance.py passes (10 tests), including the
  existing test_execution_and_range and test_alternative_relationships.
• Each new assertion fails on the unpatched source (the degenerate scalar case
  reports 1.95) and passes with the fix.
• Broad regression on the consumers of the function
  (test_moran.py, test_moran_local_mv.py) passes.

Closes #504