EEGNet on Inner Speech derivatives

First deep baseline using the same subject-wise and session-wise splits as the classical stack.

EEGNet trained on preprocessed derivatives with identical evaluation protocol. Results logged alongside CSP and Riemannian runs.