One deep music representation to rule them all? A comparative analysis of different representation learning strategies