Add padding to normal attribute in Compatibility renderer to match the RD renderers

[clayjohn]

Fixes: #83528

This is copy-pasted from the RD renderer:

godot/servers/rendering/renderer_rd/storage_rd/mesh_storage.cpp

Lines 376 to 392 in e8d57af

	// If we have an uncompressed surface that contains normals, but not tangents, we need to differentiate the array
	// from a compressed array in the shader. To do so, we allow the the normal to read 4 components out of the buffer
	// But only give it 2 components per normal. So essentially, each vertex reads the next normal in normal.zw.
	// This allows us to avoid adding a shader permutation, and avoid passing dummy tangents. Since the stride is kept small
	// this should still be a net win for bandwidth.
	// If we do this, then the last normal will read past the end of the array. So we need to pad the array with dummy data.
	if (!(new_surface.format & RS::ARRAY_FLAG_COMPRESS_ATTRIBUTES) && (new_surface.format & RS::ARRAY_FORMAT_NORMAL) && !(new_surface.format & RS::ARRAY_FORMAT_TANGENT)) {
	// Unfortunately, we need to copy the buffer, which is fine as doing a resize triggers a CoW anyway.
	Vector<uint8_t> new_vertex_data;
	new_vertex_data.resize_zeroed(new_surface.vertex_data.size() + sizeof(uint16_t) * 2);
	memcpy(new_vertex_data.ptrw(), new_surface.vertex_data.ptr(), new_surface.vertex_data.size());
	s->vertex_buffer = RD::get_singleton()->vertex_buffer_create(new_vertex_data.size(), new_vertex_data, use_as_storage);
	s->vertex_buffer_size = new_vertex_data.size();
	} else {
	s->vertex_buffer = RD::get_singleton()->vertex_buffer_create(new_surface.vertex_data.size(), new_surface.vertex_data, use_as_storage);
	s->vertex_buffer_size = new_surface.vertex_data.size();
	}

I missed adding this to the Compatibility renderer in the original PR.

Before:

After:

[akien-mga]

Since this is a copy paste from the RD renderer, I'm confident approving this.

[akien-mga]

Thanks!