Fix AVX2 SIMD boundary overread in colorShuffle_avx2

This commit fixes memory overread issues in AVX2-optimized RGB/BGR color
conversions that could cause crashes at certain image widths (e.g., 1280).

Problem:
- 3->3 conversion (RGB↔BGR): Reads 16 bytes but only needs 15, causing
  1-byte overread at boundaries
- 3->4 conversion (RGB→RGBA): Second read extends to offset+27, causing
  4-byte overread at boundaries

Root cause:
SIMD intrinsics (_mm_loadu_si128) read fixed byte amounts that can exceed
actual pixel data at image boundaries.

Solution:
Calculate proper loop boundary for each conversion type:
- 3->4 conversions: patchSize + 2 (needs 2 extra pixels margin)
- 3->3 conversions: patchSize + 1 (needs 1 extra pixel margin)
- 4->4 and 4->3: patchSize (already safe)

Testing:
Added 78 comprehensive test cases covering:
- Issue #30 exact scenario (1280x720)
- Critical boundary widths (1272-1280, 1915-1920)
- Small/tiny widths (1-10 pixels)
- Non-contiguous memory (with stride padding)

All tests pass on both CPU and AVX2 backends.

Fixes #30

Author: wysaid

src/ccap_convert_avx2.cpp

@@ -181,11 +181,21 @@ AVX2_TARGET void colorShuffle_avx2(const uint8_t* src, int srcStride, uint8_t* d
         shuffle128 = _mm_load_si128((__m128i*)shuffleData);
     }
 
+    // Different cases require different boundary conditions to avoid reading beyond allocated memory:
+    // - 3->4: reads 16 bytes from x*3+12, needs x*3+27 < width*3, i.e., x+9 < width
+    // - 3->3: reads 16 bytes from x*3, needs x*3+15 < width*3, i.e., x+5 < width  
+    // - 4->3: reads 16 bytes from x*4+16, needs x*4+31 < width*4, i.e., x+8 <= width
+    // - 4->4: reads 32 bytes from x*4, needs x*4+31 < width*4, i.e., x+8 <= width
+    constexpr uint32_t loopBoundary = (inputChannels == 3 && outputChannels == 4) ? (patchSize + 2) :
+                                      (inputChannels == 3 && outputChannels == 3) ? (patchSize + 1) :
+                                                                                    patchSize;
+
     for (int y = 0; y < height; ++y) {
         const uint8_t* srcRow = src + y * srcStride;
         uint8_t* dstRow = dst + y * dstStride;
         uint32_t x = 0;
-        while (x + patchSize <= (uint32_t)width) {
+        
+        while (x + loopBoundary <= (uint32_t)width) {
             // _mm256_shuffle_epi8 can’t move these bytes across 16-byte lanes of the vector.
             // @see issue <https://stackoverflow.com/questions/77149094/how-to-use-mm256-shuffle-epi8-to-order-elements>
             if constexpr (outputChannels == 4 && inputChannels == 3) { // 3 -> 4, need to split channels

tests/CMakeLists.txt

@@ -73,6 +73,7 @@ add_executable(
     test_yuv_conversions.cpp
     test_platform_features.cpp
     test_frame_conversions.cpp
+    test_boundary_conditions.cpp
 )
 
 target_link_libraries(