SIMD ColorLUT. First try

Two items, not two bytes

SIMD ColorLUT. remove SHIFT_ROUNDING

SIMD ColorLUT. improve performance by preliminary index calculation

SIMD ColorLUT. table_channels==4 case, minor optimizations

SIMD ColorLUT. remove unused utility

SIMD ColorLUT. remove left_mask and right_mask

SIMD ColorLUT. AVX2 implementation (near the same speed)

SIMD ColorLUT. Fast AVX2 implementation with very wired slowdown

SIMD ColorLUT. finally fix alpha copy

SIMD ColorLUT. 16bit arithmetic. Access violation fixed

src/_imaging.c

@@ -798,7 +798,7 @@ _prepare_lut_table(PyObject *table, Py_ssize_t table_size) {
     }
 
     /* malloc check ok, max is 2 * 4 * 65**3 = 2197000 */
-    prepared = (INT16 *)malloc(sizeof(INT16) * table_size);
+    prepared = (INT16 *)malloc(sizeof(INT16) * (table_size + 2));
     if (!prepared) {
         if (free_table_data) {
             free(table_data);