On Wayland, pre-multiply alpha for custom cursor

Fixes: #3360

examples/custom_cursors.rs

@@ -56,6 +56,7 @@ fn main() -> Result<(), impl std::error::Error> {
     let custom_cursors = [
         decode_cursor(include_bytes!("data/cross.png"), &event_loop),
         decode_cursor(include_bytes!("data/cross2.png"), &event_loop),
+        decode_cursor(include_bytes!("data/gradient.png"), &event_loop),
     ];
 
     event_loop.run(move |event, _elwt| match event {
@@ -72,7 +73,7 @@ fn main() -> Result<(), impl std::error::Error> {
                 Key::Character("1") => {
                     log::debug!("Setting cursor to {:?}", cursor_idx);
                     window.set_cursor(custom_cursors[cursor_idx].clone());
-                    cursor_idx = (cursor_idx + 1) % 2;
+                    cursor_idx = (cursor_idx + 1) % 3;
                 }
                 Key::Character("2") => {
                     log::debug!("Setting cursor icon to default");

src/platform_impl/linux/wayland/types/cursor.rs

@@ -37,12 +37,15 @@ impl CustomCursor {
             )
             .unwrap();
 
-        for (canvas_chunk, rgba_chunk) in canvas.chunks_exact_mut(4).zip(image.rgba.chunks_exact(4))
-        {
-            canvas_chunk[0] = rgba_chunk[2];
-            canvas_chunk[1] = rgba_chunk[1];
-            canvas_chunk[2] = rgba_chunk[0];
-            canvas_chunk[3] = rgba_chunk[3];
+        for (canvas_chunk, rgba) in canvas.chunks_exact_mut(4).zip(image.rgba.chunks_exact(4)) {
+            // Alpha in buffer is premultiplied.
+            let alpha = rgba[3] as f32 / 255.;
+            let r = (rgba[0] as f32 * alpha) as u32;
+            let g = (rgba[1] as f32 * alpha) as u32;
+            let b = (rgba[2] as f32 * alpha) as u32;
+            let color = ((rgba[3] as u32) << 24) + (r << 16) + (g << 8) + b;
+            let array: &mut [u8; 4] = canvas_chunk.try_into().unwrap();
+            *array = color.to_le_bytes();
         }
 
         CustomCursor {