Fix(device_runner): add host-side stream sync timeout watchdog with best-effort reset

[indigo1973]

Problem: rtStreamSynchronize may block indefinitely when device execution stalls, leaving host-side execution without a timeout recovery path.

Solution: run rtStreamSynchronize in a worker thread and monitor it with a condvar-based watchdog (1s timeout via PLATFORM_STREAM_SYNC_TIMEOUT_MS).

Timeout handling:

Log diagnostics (stream, device_id, block_dim, worker_count) Attempt aclrtResetDevice to unblock the stuck sync Wait one grace window for worker exit
If still blocked, detach worker to avoid permanent host hang On reset success, invalidate stream pointers and acl_ready_; on reset-fail+detach, clear stream pointers to avoid finalize-time UAF (with leak risk) In pto_runtime_c_api, detect timeout via last_run_timed_out() and trigger DeviceRunner::finalize() for best-effort cleanup Files:

platform_config.h: add PLATFORM_STREAM_SYNC_TIMEOUT_MS
device_runner.h/.cpp: add synchronize_stream_with_timeout() and last_run_timed_out_
pto_runtime_c_api.cpp: add timeout-triggered finalize path

[gemini-code-assist]

Code Review

This pull request implements a host-side watchdog for stream synchronization to prevent permanent hangs. It introduces the synchronize_stream_with_timeout method, which executes synchronization in a separate thread and handles timeouts by resetting the device or detaching the worker thread. A review comment identifies a potential resource leak where clearing the acl_ready_ flag prematurely prevents aclFinalize() from being called during the subsequent cleanup process.