Combining EEGNet with SPDNet Towards an End-To-End Architecture for Imagined Speech Decoding