arm: avoid overflow in RGBA palette offset

[soul-sol]

Summary

Promote row_width to size_t before multiplying it by the four-byte RGBA pixel size in the ARM NEON palette expansion path.

Without the promotion, 4 * row_width is evaluated as 32-bit unsigned arithmetic and wraps for widths greater than UINT32_MAX / 4. This matches the existing ARM NEON RGB path and the scalar RGBA implementation.

Fixes #862.

Verification

• Confirmed the original expression wraps to offset 3 for row_width = 1,073,741,825, while the promoted expression yields 4,294,967,299.
• Compiled pngsimd.c with Clang as an ARM64 Mach-O object using -Wall -Wextra -Werror.
• Built shared and static libpng for AArch64 Ubuntu 24.04 with CMake.
• Ran the full CTest suite on AArch64: 37/37 tests passed.
• git diff --check passes.

[jbowler]

Hum. As I implied in the comments I just attempted to preserve the "original code" bugs and all. At this point the code requires a maintainer.

Maintainer, step forward.