Avoid HF Hub access in CPU unit test setup (#8053)
The `cpu-torch-latest / unit tests` lane can fail with Hugging Face Hub
HTTP 429 errors during test collection/setup, before it reaches
PR-specific assertions
([example](https://github.com/deepspeedai/DeepSpeed/actions/runs/27073376531/job/79997387234?pr=8038)).
This is existing test flakiness rather. DeepSpeed has already had
similar 429 failures in the inference tests (#5715), and the nested ZeRO
init test also relies on a remote Hugging Face tiny model path added in
#3886.
This PR removes live Hub access from those unit-test setup paths. It
encodes the valid inference model/task combinations locally instead of
calling `HfApi().list_models(...)`, and it creates a tiny local
`VisionEncoderDecoderModel` artifact so the test can still exercise
`from_pretrained(...)` without downloading from the Hub.
Signed-off-by: Masahiro Tanaka <mtanaka@anyscale.com>
