Reductions on arrays containing `nan` give unexpected results

[majosm]

Describe the bug Array reductions such as min and max produce unexpected results when nans are present in the input array.

To Reproduce

import pyopencl as cl
import pyopencl.array as cla
import numpy as np

ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)

a = np.full(10, np.nan)
a_dev = cla.to_device(queue, a)

print(f"{cla.max(a_dev)=}")

b = np.full(10, np.nan)
b[0] = 1.
b_dev = cla.to_device(queue, b)

print(f"{cla.max(b_dev)=}")

Output:

cla.max(a_dev)=cl.Array(-inf)
cla.max(b_dev)=cl.Array(1.)

Expected behavior Ideally these should both return cl.Array(nan)?

Environment (please complete the following information):

• OS: macOS Catalina
• ICD Loader and version: whatever macOS uses (khronos-opencl-icd-loader?)
• ICD and version: pocl 1.8
• CPU/GPU: Intel i7
  • Python version: 3.9
  • PyOpenCL version: 2022.1.6

[inducer]

Could you start a PR with a failing regression test to save me time?

[majosm]

Could you start a PR with a failing regression test to save me time?

https://github.com/inducer/pyopencl/pull/584

minimum/maximum also appear to have some nan-handling issues.