add test-time dependency on setuptools

[jameslamb]

Summary

Contributes to #775, follow-up to #773.

Some nightly Python 3.12 tests are failing in this repo, as a result of the following:

• some tests here rely on GPUtil
• GPUtil was last updated 5+ years ago, and unconditionally imports distutils (code link)
• distutils was removed from the Python standard library in Python 3.12

This proposes a temporary fix... adding a test-time dependency on setuptools. import distutils will work in an environment where setuptools is installed.

Benefits of these changes

Fixes nightly tests, without imposing any new requirements of users of cucim.

Notes for Reviewers

At some point, setuptools will likely remove distutils or the parts of it that GPUtil needs, and these tests will break in this way again. I've listed out some alternatives in #775, those should be explored at some point.

How I tested this

On a machine with CUDA 12.2 and V100s, I can reproduce the test failures like this:

docker run \
    --rm \
    --gpus 1 \
    -v $(pwd):/opt/work \
    -w /opt/work \
    -it rapidsai/citestwheel:cuda12.5.1-ubuntu22.04-py3.12 \
    bash

python -m pip install \
    'cucim-cu12[test]==24.10.*,>=0.0.0a0'

python -m pytest \
    --numprocesses=8 \
    --dist=worksteal \
    ./python/cucim/tests/performance/clara

After running pip install setuptools, those tests pass 😁

[jameslamb]

setuptools==60.0.0 is the first release where disutils was vendored. From https://setuptools.pypa.io/en/latest/deprecated/distutils-legacy.html

Since the 60.0.0 release, Setuptools includes a local, vendored copy of distutils (from late copies of CPython) that is enabled by default.

That release is from December 2021 (link), so I'd be surprised if it causes any conflicts.

[bdice]

We may be able to get rid of GPUtil. It's very outdated and we have other ways to get the information it provides.

It is only used in one test:

cucim/python/cucim/tests/performance/clara/test_read_region_memory_usage.py

Line 24 in 1584885

def test_read_region_cuda_memleak(testimg_tiff_stripe_4096x4096_256_jpeg):

We can replace GPUtil.get_GPUs()[0].memoryUsed with this code from rmm:

import rmm

def get_used_memory():
    free, total = rmm.mr.available_device_memory()
    memory_used = (total - free) / (2 ** 20)
    return memory_used

Or we can use cuda-python: https://nvidia.github.io/cuda-python/module/cudart.html#cuda.cudart.cudaMemGetInfo

import cuda.cudart

def get_used_memory():
    status, free, total = cuda.cudart.cudaMemGetInfo()
    if status != cuda.cudart.cudaError_t.cudaSuccess:
        raise RuntimeError("Failed to get GPU memory info.")
    memory_used = (total - free) / (2 ** 20)
    return memory_used

Neither rmm or cuda-python are test dependencies at present, but I think it would be reasonable to use one of these stable, actively maintained packages instead of GPUtil. I lean towards the cuda-python solution because it's more minimal.