Add AVX SIMD implementations of video_orc_chroma_down_h2_u8 and video_orc_chroma_down_v2_u8

[Copilot]

With ORC disabled, video_orc_chroma_down_h2_u8 and video_orc_chroma_down_v2_u8 fall back to scalar C loops processing one element at a time. This adds AVX1 vectorized paths that process 32 bytes per iteration.

Approach

Both functions reduce to: average specific UV bytes, keep other bytes unchanged. This maps cleanly to VPAVGB + VPBLENDVB — average everything, then blend to select only the chroma bytes from the averaged result.

• h2 (horizontal): UV bytes from the second pixel in each pair need to be shuffled (VPSHUFB) to align with the first pixel's UV position before averaging
• v2 (vertical): Two source rows are already aligned — just VPAVGB the rows and blend

Implementation details

• AVX1 only — VMOVDQU (256-bit load/store), VEXTRACTF128/VINSERTF128 (lane split/merge), VPSHUFB, VPAVGB, VPBLENDVB (128-bit VEX-encoded SSE)
• Three-tier loop: 32-byte AVX → 16-byte SSE → scalar remainder
• Runtime dispatch via __builtin_cpu_supports("avx"), falls back to original scalar code on non-AVX CPUs
• Guarded by DISABLE_ORC && (__GNUC__ || __clang__) && (x86_64 || i386)
• Original scalar code is completely preserved as-is

Core pattern (v2 shown)

// Average all bytes, then blend to keep AY from s1 and averaged UV
__m128i avg = _mm_avg_epu8(s1_lo, s2_lo);
__m128i result = _mm_blendv_epi8(s1_lo, avg, blend_mask);
//                                ^^^^^ AY    ^^^ UV

Verification

Tested against scalar reference across 15 sizes (0–1024), edge cases, and in-place operation — bitwise identical results.