Grain optimizations

fguv_32x32xn_8bpc_420_csfl0_c: 8945.4
fguv_32x32xn_8bpc_420_csfl0_avx2: 1001.6
fguv_32x32xn_8bpc_420_csfl1_c: 6363.4
fguv_32x32xn_8bpc_420_csfl1_avx2: 1299.5
fgy_32x32xn_8bpc_c: 17997.6
fgy_32x32xn_8bpc_avx2: 2622.1
gen_grain_y_ar0_8bpc_c: 92326.1
gen_grain_y_ar0_8bpc_avx2: 20641.1
gen_grain_y_ar1_8bpc_c: 160131.6
gen_grain_y_ar1_8bpc_avx2: 67012.1
gen_grain_y_ar2_8bpc_c: 242255.6
gen_grain_y_ar2_8bpc_avx2: 109997.5
gen_grain_y_ar3_8bpc_c: 408954.9
gen_grain_y_ar3_8bpc_avx2: 119569.1