Implement `generate` (inference) for torch exported text-generation models

### Feature request Unlike `torch.compile`, `torch.export` only export the inner transformer to predict a single token for text-generation models. The autoregressive logics are not included in the exported artifact. For ExecuTorch Runtime, the autoregressive logics is implemented there, either directly in c++ ([code pointer](https://github.com/pytorch/executorch/blob/main/examples/models/llama2/runner/runner.cpp#L310-L415)) or python ([code pointer](https://github.com/pytorch/executorch/blob/main/examples/models/llama2/runner/generation.py#L70-L124)). Here is a minimal `generate` impl in a test for Phi3-mini in 🤗 ([code pointer](https://github.com/huggingface/transformers/blob/543df489147412efb20575c3c2a3fb69a18ac107/tests/models/phi3/test_modeling_phi3.py#L73)). Instead of duplicate the generate logics for each exported model, it's better to have a common implementation of `generate` that can be used by any exported text-generation mdoel. ### Motivation To support the new "Export to ExecuTorch" workflow. This would enable users to have a unified `generate` experience as they run inference using eager or compiled model. ### Your contribution TBD