add release note and up blocksort size

Author: Intron7

docs/release-notes/0.15.1.md

@@ -8,6 +8,7 @@
 ```{rubric} Bug fixes
 ```
 * Fixes `tl.rank_genes_groups` returning NaN/zero `logfoldchanges`/`pvals` with `groups=[subset]` and `reference='rest'` {pr}`651` {smaller}`S Dicks`
+* Fixes `pp.bbknn` connectivities diverging from upstream `bbknn`: per-batch neighbours are now sorted by distance before `fuzzy_simplicial_set` (so weights no longer collapse near 1.0), and the default `trim` matches upstream (`10 * neighbors_within_batch * n_batches`). Trimming kernel no longer crashes for large `trim`, and a new block-cooperative sort kernel is auto-dispatched for large `trim` for substantial speedups {pr}`659` {smaller}`S Dicks`
 
 ```{rubric} Misc
 ```

src/rapids_singlecell/_cuda/bbknn/bbknn.cu

@@ -6,10 +6,10 @@
 using namespace nb::literals;
 
 constexpr int BLOCK_SIZE = 64;
-// Block-cooperative sort kernel: BLOCK_THREADS * ITEMS_PER_THREAD = 1024.
+// Block-cooperative sort kernel: BLOCK_THREADS * ITEMS_PER_THREAD = 2048.
 // Rows larger than this must use the per-thread kernel (kernel 1).
 constexpr int SORT_BLOCK_THREADS = 128;
-constexpr int SORT_ITEMS_PER_THREAD = 8;
+constexpr int SORT_ITEMS_PER_THREAD = 16;
 constexpr int SORT_TILE_SIZE = SORT_BLOCK_THREADS * SORT_ITEMS_PER_THREAD;
 
 static inline void launch_find_top_k_per_row(const float* data,