Isolate Hugging Face dataset cache - #13

Fixes #12.

What

• Give each non---keep-cache fetch a process-owned temporary datasets cache.
• Place it inside the resolved output directory so large Arrow intermediates use the corpus filesystem rather than a potentially small /tmp.
• Pass that path explicitly to load_dataset(cache_dir=...).
• Remove only the owned path on success and conversion failure.
• Complete cleanup before the streaming os._exit(0) path.
• Fail instead of exiting successfully when cleanup fails.
• Report both conversion and cleanup errors when both occur.
• Leave --keep-cache on the normal shared Hugging Face cache without overriding it.

Why

The previous default recursively deleted ~/.cache/huggingface/datasets, which may contain unrelated datasets belonging to other projects. Mini-AGI should only remove cache files it created and owns.

Verification

python -m unittest discover -s tests -v

Seven tests pass. They cover:

• preserving an unrelated shared-cache sentinel;
• owned cache placement, use, and cleanup;
• --keep-cache using the library default;
• cleanup after load failure;
• cleanup before streaming immediate exit;
• cleanup failure preventing successful exit;
• combined conversion and cleanup error reporting;
• resolved-path cleanup after output symlink retargeting.

Also verified with compileall and git diff --check. The patch is independent of open PRs #7, #9, and #11.