Fix async durability checkpoint backlog

[Alexxigang]

Summary

Fixes unbounded checkpoint task buildup when running graphs with durability="async".

The root cause is that a new checkpoint write can be chained before the previous async checkpoint future has finished, which allows _checkpointer_put_after_previous tasks to accumulate across supersteps when checkpoint persistence is slower than graph execution.

This change keeps the intended one-superstep overlap for async durability, but waits for the in-flight checkpoint future before chaining the next one.

Changes

• wait for the pending checkpoint future before calling loop.after_tick() in the sync Pregel stream path
• wait for the pending checkpoint future before calling loop.after_tick() in the async Pregel stream path
• add regression coverage for:
  • async durability backlog stays bounded
  • async durability still preserves single-superstep overlap
  • sync durability still waits for checkpoint completion
  • exit durability still defers checkpoint work until completion
  • final async state is flushed before returning

Testing

python -m py_compile libs/langgraph/langgraph/pregel/main.py libs/langgraph/tests/test_async_durability.py
python -m pytest --noconftest libs/langgraph/tests/test_async_durability.py -q

Passed locally:

• 5 passed in 11.80s

Issue

Closes #7094