This PR implements two optimizations to coalescedReductionThinKernel which is used for coalesced reductions of tall matrices (many rows) and/or thin (few columns):

1. Process multiple rows per warp to increase bytes in flight and amortize load latencies.
2. Use a vectorized reduction to avoid the LSU bottleneck and have fewer global stores (and at least partially coalesced).

The benchmark below shows the achieved SOL percentage on A30. I also measured that on H200, it achieved 84% SOL for 32 columns and up to 94% for 512 columns.