Text this: Multimodal Explainability Using Class Activation Maps and Canonical Correlation for MI-EEG Deep Learning Classification