Fix/issue 20117 runner aenter race

[jellyfish0316]

related to #20117

this PR prevents concurrent Runner auto-starts from racing on shared async lifecycle state.

Summary

When multiple flow runs try to auto-start the same Runner concurrently, they could race in Runner.__aenter__ and attempt to enter or tear down shared async contexts more than once. In practice this could surface as:

• RuntimeError("Each CancelScope may only be used for a single 'with' block")
• RuntimeError("Attempted to exit cancel scope in a different task than it was entered in")
• Cannot open a client instance more than once

This change serializes runner lifecycle transitions with a private lock, tracks active context users, and ensures final teardown is performed by the task that entered the shared async context stack.

It also adds a focused regression test covering concurrent implicit startup through execute_bundle.

Validation:

• uv run pytest tests/runner/test_runner.py -k "test_concurrent_execute_bundle_initializes_runner_once" -vv
• local high-concurrency repro: 100 iterations of concurrent bundle execution completed without reproducing the runner lifecycle errors above

[codspeed-hq]

Merging this PR will not alter performance

✅ 2 untouched benchmarks

---

_{Comparing jellyfish0316:fix/issue-20117-runner-aenter-race (2466170) with main (98309ff)}