Commit 571683b8e795aff8314088132f571624a3d2c6bd ("Add reference OpenVINO IRs") replaced the full checkpoint with only the OpenVINO IR files and metadata.json, dropping config.json, generation_config.json, and all tokenizer/preprocessor files. This breaks `AutoConfig.from_pretrained(..., revision="ov")` since there is no `model_type` in config.json (there is no config.json at all), which in turn breaks optimum-intel's `optimum_pipeline(..., revision="ov")` / `OVModelForSpeechSeq2Seq.from_pretrained` loading path (see optimum-intel tests/openvino/test_modeling.py::PipelineTest::test_seq2seq_load_from_hub).
This PR restores the files removed in that commit (as they were in f50ad03ad4b24324d29328285a2e019d081135aa, the last commit before the reference IR files were added), keeping the newer OpenVINO IR + metadata.json files.
