fix d2d cupy transfer

[Intron7]

Fix multi-GPU cudaErrorLaunchFailure during cross-device result aggregation when using RMM without pool allocation (pool_allocator=False).

The error manifests as a MemoryError on a tiny allocation during Phase 4 (cross-device copy dev1 → dev0):

MemoryError: std::bad_alloc: CUDA error (failed to allocate 15592 bytes) at:
.../rmm/mr/cuda_memory_resource.hpp:51: cudaErrorLaunchFailure unspecified launch failure

Root cause: When CuPy's cp.asarray() performs a cross-device D2D copy and the result is consumed inline (sums+= cp.asarray(data["sums"])), the fused allocation + copy + addition exposes a stale CUDA async error through cudaMalloc (called by RMM's cuda_memory_resource). Splitting the D2D copy into a separate assignment avoids this.

The bug is hidden by:

• RMM pool allocator — never calls cudaMalloc, so the stale error is never surfaced
• Small datasets — allocations complete before the stale state propagates

Affected functions: edistance.pairwise, co_occurrence, moran, geary (all multi-GPU paths)

Fix: Split inline cp.asarray D2D copy + addition into two statements, and sync non-blocking streams instead of null stream where applicable.

Before (fails with RMM no-pool on large workloads)

pairwise_sums += cp.asarray(data["sums"])

After (works reliably)

dev0_sums = cp.asarray(data["sums"])
pairwise_sums += dev0_sums

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.50%. Comparing base (1407937) to head (22949e5).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #594   +/-   ##
=======================================
  Coverage   87.49%   87.50%           
=======================================
  Files          96       96           
  Lines        6781     6784    +3     
=======================================
+ Hits         5933     5936    +3     
  Misses        848      848           

Files with missing lines	Coverage Δ
...apids_singlecell/pertpy_gpu/_metrics/_edistance.py	91.84% <100.00%> (+0.04%)	⬆️
src/rapids_singlecell/squidpy_gpu/_co_oc.py	92.00% <100.00%> (+0.06%)	⬆️
src/rapids_singlecell/squidpy_gpu/_gearysc.py	92.68% <100.00%> (ø)
src/rapids_singlecell/squidpy_gpu/_moransi.py	91.86% <100.00%> (ø)