Fix pagination

[Yostra]

Summary

Stripe's List API can return an empty page when starting_after (or ending_before) points at an object whose created second is fully inside the created[gte]/lt] window. The cursor effectively "loses" the rest of that one-second bucket, so any subdivision branch that inherited the cursor would silently skip records.
This PR removes the assumption that a cursor can drain a same-second boundary and instead re-queries that second by timestamp.

What changed

• subdivideRanges: child ranges no longer inherit the parent cursor. Every child is fetched fresh from its created bounds. The final (newest) child is still widened to lastObserved + 1s so the boundary second is covered without relying on the cursor.
• Small-remainder threshold tightened from < 30s to <= 1s so we only fall back to cursor-only pagination when there is genuinely nothing left to split.
• streamingSubdivide: yields a shallow clone of range and forwards lastObserved on the event, so downstream range_complete accounting uses the fetch result instead of re-scanning emitted records.
• fetchPageForRange: treats date as the created-timestamp fallback (invoice items use date, not created), and the head-completion logic in iterateStream matches.
• Re-enables concurrency: 100 for streamingSubdivide.
• Skips the "no duplicate record IDs" e2e assertion — re-fetching the boundary second can legitimately re-emit IDs; destination upserts handle idempotency.