[shortfin llm integration tests] Add tinystories llama2 25m model as a smoke test

[renxida]

This is a tiny model that would take less than a minute to download / export / compile.

The smoke test is intended to run before we run the full integration CI suite. This would give us early visibility into shortfin failures & save CI resources.

This PR adds an integration_tests function to:

• download hf repo with model.safetensors
• convert to gguf
• convert to irpa

Closes #933

[renxida]

Comments addressed. Re-review pls.

Coming next in follow-up prs:

• dedupe more code between here and sharktank's hf_datasets.py
• maybe move model_management and related files into shark-ai repo. Not sure what's the plan / if it's ok to make it less of a stub