Re: [ot][spam][crazy][data] transformer model 'attention' improvement

- torch tensors do not allocate new memory when expanded, and are documented as views that have that property next: jax tensors, and model logic