u16 blend operations are 2x slower than Skia. No idea why. Maybe we should try using u16x8 lane instead of u16x16 for ARM
rectangle fill is like 6x slower during benchmarks for no reason, but only 2x slower when run separately
gradients are 2-3x slower, probably because we have to use f32x16, and the current f32<->u16 cast code doesn't use NEON. f32/highp is actually faster, which is bizarre
Currently we're 2-4x slower in some benchmarks on Apple M1. This is with ARM NEON.
Benchmark results: https://razrfalcon.github.io/tiny-skia/arm.html
Otherwise, performance is decent.