fix(workflow): add concurrency control and fix timer race conditions

[KhawarHabibKhan]

Description

• Add asyncio.Lock to guard step_n shared state in the workflow engine, preventing race conditions when step_semaphore > 1
• Fix TOCTOU race in RDAgentTimer.is_timeout() by using cached _remain_time_duration instead of calling datetime.now() twice
• Clamp _remain_time_duration to non-negative to prevent negative timer values propagating through the system
• Add defensive clamp on MLflow remain_time and remain_percent metrics
• Update DataScienceRDLoop._check_exit_conditions_on_step subclass override to match new async signature

Motivation and Context

When step_semaphore is configured > 1, multiple async steps can concurrently read/decrement step_n without synchronization, leading to over-execution or missed termination. The timer's is_timeout() also had a TOCTOU race where update_remain_time() and a separate datetime.now() comparison could disagree, and negative _remain_time_duration values could propagate into MLflow metrics.

Types of changes

• Fix bugs
• Add new feature

---

📚 Documentation preview 📚: https://RDAgent--1347.org.readthedocs.build/en/1347/