docs: explain offline model loading

[widechaos]

Users running on offline machines or unreliable networks need to know how to reuse model files without contacting the Hugging Face Hub. Add README examples for downloading a portable local model directory and loading an existing cache with local_files_only=True.

Explain custom cache directories and the tokenizer fallback: a local model still needs tokenizer.json, otherwise loading may attempt a separate Hub request. This complements the offline deployment question in #1430.

Validation: all 19 existing tests passed; Black, isort, and Flake8 checks passed. Parsed the new Python examples and verified that the cache-only argument and tokenizer allow-list are passed to the Hub download API.

[Purfview]

Could you post your guides in the "Show and Tell" category?

Alternatively, you could answer the questions directly in the threads where they were asked.