PixIT: Joint Training of Speaker Diarization and Speech Separation from Real-world Multi-speaker Recordings