x86: Rewrite SGR AVX2 asm

Checkasm numbers on a 9700K:

                    old -> new
sgr_3x3_8bpc_avx2:  60306 47152
sgr_5x5_8bpc_avx2:  46547 35244
sgr_mix_8bpc_avx2: 109326 74386

Also reduces stack usage by >200 KiB.