Fix transform when using cuML HDBSCAN

[beckernick]

This PR:

• Uses cuML's membership_vector functionality in the transform pipeline rather than approximate predict. This now matches the default CPU-based HDBSCAN behavior
• Updates the test suite to include an optional cuML-based test. Because pytest fixtures are materialized when first requested by a test runtime, we can use cuML functionality within the fixture and mark the pytest to only run in environments/systems in which cuML is available.

Closes #1764 and closes #1317

For completeness, I've included the output of a local run of this pytest using RAPIDS/cuML 24.04:

rapids-24.04-bertopic) nicholasb@nicholasb-HP-Z8-G4-Workstation:~/NVIDIA/BERTopic/tests$ pytest test_bertopic.py -v
========================================== test session starts ==========================================
platform linux -- Python 3.10.14, pytest-8.2.0, pluggy-1.5.0 -- /home/nicholasb/miniconda3/envs/rapids-24.04-bertopic/bin/python3.10
cachedir: .pytest_cache
rootdir: /home/nicholasb/NVIDIA/BERTopic
plugins: anyio-4.3.0
collected 10 items                                                                                      

test_bertopic.py::test_full_model[base_topic_model] PASSED                                        [ 10%]
test_bertopic.py::test_full_model[kmeans_pca_topic_model] PASSED                                  [ 20%]
test_bertopic.py::test_full_model[custom_topic_model] PASSED                                      [ 30%]
test_bertopic.py::test_full_model[merged_topic_model] PASSED                                      [ 40%]
test_bertopic.py::test_full_model[reduced_topic_model] PASSED                                     [ 50%]
test_bertopic.py::test_full_model[online_topic_model] PASSED                                      [ 60%]
test_bertopic.py::test_full_model[supervised_topic_model] PASSED                                  [ 70%]
test_bertopic.py::test_full_model[representation_topic_model] PASSED                              [ 80%]
test_bertopic.py::test_full_model[zeroshot_topic_model] PASSED                                    [ 90%]
test_bertopic.py::test_full_model[cuml_base_topic_model] PASSED                                   [100%]

=========================================== warnings summary ============================================
...
...
============================== 10 passed, 27 warnings in 108.42s (0:01:48) ==============================

And local run of this pytest in an environment without cuML:

(bertopic) nicholasb@nicholasb-HP-Z8-G4-Workstation:~/NVIDIA/BERTopic/tests$ pytest test_bertopic.py -v
========================================== test session starts ==========================================
platform linux -- Python 3.10.14, pytest-8.2.0, pluggy-1.5.0 -- /home/nicholasb/miniconda3/envs/bertopic/bin/python3.10
cachedir: .pytest_cache
rootdir: /home/nicholasb/NVIDIA/BERTopic
plugins: anyio-4.3.0
collected 10 items                                                                                      

test_bertopic.py::test_full_model[base_topic_model] PASSED                                        [ 10%]
test_bertopic.py::test_full_model[kmeans_pca_topic_model] PASSED                                  [ 20%]
test_bertopic.py::test_full_model[custom_topic_model] PASSED                                      [ 30%]
test_bertopic.py::test_full_model[merged_topic_model] PASSED                                      [ 40%]
test_bertopic.py::test_full_model[reduced_topic_model] PASSED                                     [ 50%]
test_bertopic.py::test_full_model[online_topic_model] PASSED                                      [ 60%]
test_bertopic.py::test_full_model[supervised_topic_model] PASSED                                  [ 70%]
test_bertopic.py::test_full_model[representation_topic_model] PASSED                              [ 80%]
test_bertopic.py::test_full_model[zeroshot_topic_model] PASSED                                    [ 90%]
test_bertopic.py::test_full_model[cuml_base_topic_model] SKIPPED (cuML not available)             [100%]

=========================================== warnings summary ============================================
...
...
========================== 9 passed, 1 skipped, 8 warnings in 97.41s (0:01:37) ==========================

[MaartenGr]

Hi @beckernick!

Thanks for the PR and your work on this. Definitely helps to have somebody from the RAPIDS team on this! I just went through your code and it all looks good to me. When the pipeline runs without any problems (which I assume it will), I'll go ahead and merge it.