Fix eval port for pre-to_edge evaluation (#2519)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/2519

2 Changes to fix eval_llama

1) max_seq_length wasn't being propogated properly. This diff plumbs it through
2) _model_call via model forward was unnecessarily complicated and didn't match ET input shape. This diff simplifies this

Reviewed By: jerryzh168

Differential Revision: D55095280

fbshipit-source-id: d056b04a89c44757ccad5eeeb454b23c92db8d4e