v4

Fix rounding.

v3

• fix a tiny memory leak
• fix broken dependency tracking for api4 VS (affects vspipe --filter-time and vspipe -g output)
• update Clang/LLVM to 15.

v2

• Add support for 1-15 bit integer format

v1

AVX2-optimized implementation of std.BoxBlur (=> box.Blur)

If AVX2 is not available, it will fall back to std.BoxBlur.

Zen3 users may not benefit from the current implementation. The measured speedup is 1.3x on skylake and icelake, and 3x on zen2.