Format Eval output and enabled cuda support (#2569)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/2569

When using eval_llama_lib, it is much faster with cuda enabled if possible. This diff enables this

In addition it wraps the output format of eval to more digestable

Reviewed By: jerryzh168

Differential Revision: D55208754

fbshipit-source-id: 8744d58064b6bcab5567a62bb2bf99fe69507aa1