Fix submission intake reliability

[isPANN]

What changed

• snapshot R2 incoming/ keys and transition only objects from that batch
• quarantine permanent input failures while retaining retryable failures
• serialize scoring workflow runs and scope R2 credentials to storage steps
• enforce the official schema, pinned library commit, and clean-run requirement
• preserve submitter and run metadata through private scoring
• validate request byte size in the Worker and stop exposing internal R2 keys

Why

The previous recursive archive could move submissions that arrived during scoring without
processing them, while one malformed object could jam the entire queue. Scored detail also
dropped metadata needed by downstream leaderboard entries.

User impact

The CLI contract is unchanged. Internal submitters still use
python -m benchmark.submit; test submissions remain private and excluded from the public
leaderboard.

Validation

• uv run pytest -q — 218 passed, 4 skipped
• wrangler deploy --dry-run
• Python compile check
• JSON/YAML parse checks
• git diff --check