Replace nalgebra with glam

`glam` compiles much faster and leverages SIMD nicely.

wgpu/src/image.rs

@@ -222,11 +222,9 @@ impl Pipeline {
         bounds: Rectangle<u32>,
         target: &wgpu::TextureView,
     ) {
-        let matrix: [f32; 16] = transformation.into();
-
         let transform_buffer = device
             .create_buffer_mapped(16, wgpu::BufferUsage::COPY_SRC)
-            .fill_from_slice(&matrix[..]);
+            .fill_from_slice(transformation.as_ref());
 
         encoder.copy_buffer_to_buffer(
             &transform_buffer,