Eagle-E

**[Edit:** Training the model on the Shepard Metzler dataset for 300k iterations should be enough. These are the performances I remember getting: - 12 layers and `shared_core=false`: - +/- 3it/s...