Fix integer power opcodes for negative exponents

[raashish1601]

For an integer base ** n where n comes from an array and is negative, the int and long long power opcodes compute 1 / base. Two problems with that:

• 0 ** -1 divides by zero. That's undefined behaviour in C. With the Windows wheel (numexpr 2.14.2 from PyPI) it kills the process with STATUS_INTEGER_DIVIDE_BY_ZERO; GCC happens to compile 1 / x without a division, so Linux builds return 0.
• The exponent is ignored, so (-1) ** -2 gives -1 instead of 1.

import numpy as np, numexpr as ne
base = np.array([-1, 0], dtype=np.int32); n = np.array([-2, -1], dtype=np.int32)
ne.evaluate("base ** n")  # Windows: process crashes; Linux: [-1, 0]

With this change a negative exponent gives base for base == ±1 and an odd exponent, 1 for base == ±1 and an even exponent, and 0 otherwise (including 0 ** -n, which keeps the result Linux builds already return, similar to integer division by zero giving 0). The scalar-exponent case is unchanged, since pow_op() already rejects negative integer constants.

Added test_negative_power_array for int32 and int64; it fails on main and passes with this change. Built the extension in a python:3.12 Docker image and ran numexpr.tests.test_numexpr (6043 passed). Also added a line to the release notes.