Use checkpoint dtype to export llama (#1870)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/1870

Inspect the state_dict to get the dtype of the checkpoint.

Assumes that checkpoint only has one dtype, otherwise prints the mismatches.

add `--dtype-override` flag to explicitly specificy fp16, fp32. Remove -H flag.

Reviewed By: larryliu0820

Differential Revision: D53505038

fbshipit-source-id: 1073a155dcdc9c390981284094e3ab9be2c3b12d