Use NVIDIA's RAPIDS cuML variant of UMAP

[JohnTigue]

There is a GPU-accelerated version of UMAP provided by the RAPIDS cuML library. The original implemention, umap-learn, is CPU only, says the docs.

(Confirm this. Numba supposedly can use CUDA GPUs, so why not?)

Some docs:

• Aug 27, 2024 RAPIDS 24.08: Better scalability, performance, and CPU/GPU interoperability
• Oct 31, 2024 Even Faster and More Scalable UMAP on the GPU with RAPIDS cuML