fix(demo): release materialization claims after rollback

[suguanYang]

Summary

• capture demo materialization claim IDs before publication begins
• release claims by stable IDs after a rollback instead of reading expired ORM instances
• preserve the original publication exception rather than masking it with SQLAlchemy MissingGreenlet

Context

A production demo-spacex-s1 materialization exceeded the database command timeout. The rollback expired the in-memory claim objects, and failure cleanup then read claim.id, causing SQLAlchemy to attempt async I/O outside greenlet_spawn. That secondary MissingGreenlet hid the original timeout and could leave the materialization claim until stale-claim recovery.

This PR fixes the failure cleanup only. It does not change the materialization database timeout or the bulk publication workload.

Validation

• uv run pytest apps/api/tests/contract/test_demo_documents_contract.py -q — 11 passed
• uv run ruff check apps/api/app/services/demo/source_materializer.py apps/api/tests/contract/test_demo_documents_contract.py
• uv run pyright apps/api/app/services/demo/source_materializer.py apps/api/tests/contract/test_demo_documents_contract.py
• git diff --check