Llama.cpp b10984: target CUDA row-contiguous reductions in upgrade testing

Treat llama.cpp b10984 as a targeted CUDA regression candidate, not a blanket inference upgrade

If your CUDA-backed runtime uses row-contiguous reductions, b10984 is worth a focused regression pass. The release description calls out CUDA support for row-contiguous SUM_ROWS, adds row-contiguous MEAN support through the same shared kernel, and mentions a MEAN permute/slice test.

That evidence supports a narrow test plan: exercise the affected operations with row-contiguous tensors, then include the permute/slice shape used by the stated MEAN test. Compare numerical outputs with the version you currently run, and record failures by operation and layout. This is a better upgrade decision than treating the release label as proof of a broad speedup.

Keep the conclusion bounded. The release provides no benchmark, end-to-end latency measurement, comprehensive correctness result, or production-reliability evidence. A passing targeted test would justify confidence for the covered paths; it would not establish that every model benefits, that other backends change, or that production is ready without workload-specific validation.