Cache harvested shards as GitHub release assets

publish_shards.sh incrementally packs not-yet-published shards into
<=1.8 GiB tarballs on the shards-v1 release, with a published.txt
index updated only after each successful upload (interruption-safe,
safe to run mid-harvest) and the provenance ledger alongside.
fetch_shards.sh is the fast path for training boxes: one tarball
in flight at a time, so peak extra disk stays ~1.8 GiB.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>