fix: make store-based barrier reusable across updates

[RunningLeon]

fix: make the shared-store barrier reusable across updates in #106

Summary

• add an independent generation counter for ParameterServer.store_based_barrier();
• use a unique _store_based_barrier() group name for every invocation;
• keep the long-lived root TCPStore unchanged and shared;
• add CPU-only regression coverage for two consecutive barriers on the same TCPStore.

[koriyoshi2041]

Verified the repeated-barrier path on the exact head. The generation counter is independent of process-group creation, every rank deterministically advances it once per successful update barrier (including auto_pg=False), and the shared-store regression exercises two consecutive generations. Locally, tests/test_store_barrier.py passes 2/2; focused Ruff check/format and diff hygiene also pass.