Preserve signed-zero branch sides in complex square roots

[antonsoo]

sqrt chooses the wrong side of the complex branch cut after conjugation:

x = np.array([-4 + 0j])
ne.evaluate("sqrt(conj(x))")  # currently +2j; expected -2j

Preserve the imaginary component's sign with copysign and return a nonnegative real zero for zero inputs. This matches NumPy and cmath.sqrt, including negative imaginary zero. Ordinary complex inputs keep their existing arithmetic.

The regression covers signed zeros, quadrants, strides, block-sized arrays, nested conjugation and aliased output. In a spectral image calculation, the corrected adjoint matches the conjugated forward operator across 18 configurations; before, 205,254 frequency-bin results differed. Ordinary complex controls remain byte-identical.

Validation: the new regression fails before and passes after; the full installed native suite passes 6,042 tests on Python 3.12 / NumPy 2.5.4, without VML. Applicable pre-commit hooks pass.