Fix the uint8 wrap model for pixel normalization

An AI normalized uint8 image pixels for a CNN, subtracting an integer mean while the array was still uint8. In unsigned 8-bit arithmetic, a pixel below the mean wraps around to a large value instead of going negative, with no error, and training stalled on a plateau. Your team models the normalization both ways to compare them.

An assistant wrote normalize_pixel_batch(pixels, mean_value, scale, wrap_uint8). It is in the editor. It has a bug. Find it and fix it.

Input and output:

• pixels: pixel values from 0 to 255.
• mean_value: the mean to subtract.
• scale: the number to divide by.
• wrap_uint8: true to model the uint8 subtraction, false for the float path.
• Return the normalized values, in order.

Rules the tests check:

• If scale is 0, return an empty list.
• For each pixel, take the difference: the pixel minus mean_value.
• When wrap_uint8 is true, replace the difference with the difference modulo 256, a whole number from 0 to 255, the way uint8 subtraction wraps. So 5 minus 10 becomes 251.
• Then divide by scale.
• Answers within 0.000001 of the expected values pass.