Use the float32 kernel for nextafter on float32 inputs

[raashish1601]

nextafter is registered with a minimum kind of double, so float32 inputs are upcast and the step is taken in float64. The result is a float64 array whose values are one float64 ulp away from the inputs, not the next float32 values:

import numpy as np, numexpr as ne
x = np.array([1.0, 0.0], dtype=np.float32); y = np.array([2.0, 1.0], dtype=np.float32)
ne.evaluate("nextafter(x, y)")  # array([1.0e+000, 4.9e-324])  (float64)
np.nextafter(x, y)              # array([1.0000001e+00, 1.4012985e-45], dtype=float32)

The nextafter_fff kernel already exists in functions.hpp (and in the MSVC stubs), so this registers nextafter with a minimum kind of float, like fmod and arctan2. Int inputs still go to double.

hypot, copysign, maximum and minimum are also registered with double even though float32 kernels exist. Their values are right there, only the result dtype is float64 instead of float32, so I left them alone. Happy to switch them too if you want.

Added test_nextafter_float32, which checks the dtype and the exact values against NumPy. It fails on master and passes with this change. Built in a python:3.12 Docker image and ran numexpr.tests.test_numexpr (6043 passed). Also added a line to the release notes.