fix: prune stale node identity state on timeout

[abendrothj]

Summary

• remove timed-out nodes from node_identities so ghost peers cannot block RDMA placement
• remove the same node from the newer per-node node_backends map
• preserve identity and backend state for live nodes
• add regression coverage for the timeout transition

Testing

• uv run pytest src/exo/shared/tests/test_apply/test_apply_node_timed_out.py
• uv run basedpyright src/exo/shared/apply.py src/exo/shared/tests/test_apply/test_apply_node_timed_out.py
• uv run ruff check src/exo/shared/apply.py src/exo/shared/tests/test_apply/test_apply_node_timed_out.py

Fixes #2194

---

Refiled from #2211, which was closed unintentionally.