Wilcoxon rank-sum addition to rank_genes_groups

[EriJFle]

Hi,

This work introduces a CuPy-based GPU implementation of the Wilcoxon rank-sum test for rank_genes_groups().

Key changes

GPU acceleration: Rank computation, group rank sums, z-scores, and two-sided p-values are computed on the GPU (CuPy + cupyx.scipy.special.ndtr).

Vectorized group operations: Replace per-group loops with a single matrix multiply group_matrix.T @ ranks to obtain all group rank sums at once.

GPU-native rank computation: Mid-ranks and tie-correction implemented via CuPy primitives (cp.argsort, cp.cumsum) to mirror Scanpy semantics.

Dynamic GPU memory management: _choose_chunk_size() queries cp.cuda.runtime.memGetInfo() to size gene chunks adaptively (avoids OOM and maximizes throughput).

Testing:
Added tests/test_rank_genes_groups_wilcoxon.py to ensure same output as Scanpy’s wilcoxon rank_genes_groups.

[asmlgkj]

when will this function merged, thanks a lot, very needed

[asmlgkj]

does it mean it fully support the Wilcoxon rank-sum on gpu?
thanks a lot @EriJFle
how I can update this

[Intron7]

@EriJFle great work on this. #535 is the PR that will be merged. It's based on your branch so your initial contributions will still be in the codebase. I also added your name to the release note. I made this even faster and more memory efficient on CPU and GPU. I also the option to define a reference.