pyucell 0.7.3 on Python PyPI

New in this version

Added GPU support for large datasets by @earmingol in #4
New optional install extra: pip install pyucell[gpu] (adds torch>=2.1).
Vectorised neighbour-weight construction in smooth_knn_scores: the per-cell
Python loop is replaced by a single sparse matmul, faster even on CPU.
Implement scoring from pre-computed rank matrix. This can be useful to test
new signature without recalculating ranks, but can demand more memory as it
requires storing the full rank matrix (unlike the regular pipeline, which
utilizes chunks of cells).
Fixed some edge cases, e.g. a 1-gene signature with max_rank=1 and
missing_genes="skip" and all signature genes are absent.
Fixed smooth_knn_scores silently overwriting adata.obsp["connectivities"]
when called without graph_key. The internally computed graph is now stored
under the dedicated key "smooth_knn_connectivities" to avoid clobbering
pre-existing neighbor graphs used for UMAP/clustering.
Fixed self-loop double-counting in _build_weight_matrix: diagonal entries
are now stripped from the input graph before computing neighbor weights.

Full Changelog: v0.6.0...v0.7.3