Fix numpy 2.5 incompatibility in Venn-Abers calibration

[JayeshSuryavanshi]

Problem

hiclass is broken on numpy 2.5. numpy 2.5 removed the 2-D form of np.cross (np.cross now requires 3-dimensional vectors), and _InductiveVennAbersCalibrator uses it on 2-D points in its convex-hull orientation checks:

def _non_left_angle_turn(self, next_to_top, top, p_i):
    res = np.cross((top - next_to_top), (p_i - top))
    return res <= 0

On numpy 2.5 this raises:

ValueError: Both input arrays must be (arrays of) 3-dimensional vectors, but they are 2 and 2 dimensional instead.

which breaks all Venn-Abers calibration. On a clean checkout with numpy 2.5.1, 14 tests fail, every calibration test that exercises the ivap/cvap methods (across test_calibration.py, test_LocalClassifierPerNode.py, test_LocalClassifierPerLevel.py, test_LocalClassifierPerParentNode.py).

Fix

For 2-D vectors a, b, np.cross(a, b) returns the scalar a[0] * b[1] - a[1] * b[0] (the signed area / orientation). Replace the np.cross call with that explicit expression, which is identical to the old behavior and works on every numpy version:

def _non_left_angle_turn(self, next_to_top, top, p_i):
    a = top - next_to_top
    b = p_i - top
    res = a[0] * b[1] - a[1] * b[0]
    return res <= 0

Tests

With this change the full suite goes from 14 failed / 385 passed to 399 passed / 0 failed on numpy 2.5.1, with no regression. I also added test_venn_abers_angle_turn_orientation, a small direct test of the two orientation helpers, so the sign convention is locked independently of the numpy version.

black --check and flake8 are clean on the changed files.