ArrowMetal 0.4.0 brings DataFusion sorts to Apple Silicon GPU, up to 28.8x faster

The datafusion-arrowmetal 0.4.0 Rust crate adds one physical optimizer rule that offloads DataFusion sorts to the Apple Silicon GPU. Full sorts of 250,000 to 50,000,000 rows run 6.9x to 28.8x faster on an M4 Max, with zero mismatches across 13,632 query pairs.

• 50M-row sort: 126 CPU-ms vs DataFusion's 6,104
• 6.9x to 28.8x faster on Apple M4 Max, 16 partitions
• 13,632 query pairs, rule off vs on: 0 mismatches
• Top-k and filters lag (0.14x to 0.79x) and are off by default