Fix discrepancy on eos and bos tokens for different pte files (#3923)

Summary:
Currently the logic is that if pte file doesn't contain metadata such as vocab size, bos id and eos id, use a default value. This doesn't work well because llama2 and llama3 are using different tokenizer artifacts, therefore they have different default values.

Instead assuming a default value at the runner level, we should deduce the values for vocab size bos id and eos id from the tokenizer aritifact, and let the metadata in the pte file override it.

This PR fixes this issue.

Pull Request resolved: https://github.com/pytorch/executorch/pull/3923

Test Plan:
demo app works for pte file without vocab size and eos id bos id:

![Simulator Screenshot - iPhone 15 Pro - 2024-06-10 at 18 07 21](https://github.com/pytorch/executorch/assets/8188269/923c60c5-0bec-4c4b-879f-419e04e52481)

Reviewed By: shoumikhin

Differential Revision: D58393326

Pulled By: larryliu0820

fbshipit-source-id: e306493c569d2350fe6df6617200c337c627024c