Make max_tasks_per_child configurable to fix CUDA memory leaks

[yuxuan-z19]

• Problem: PyTorch + multiprocessing with spawn avoids “Cannot re-initialize CUDA in forked subprocess,” but long-lived workers hold CUDA contexts, causing GPU memory to accumulate.
• Solution: Added max_tasks_per_child as a configurable option in the executor. Setting it (e.g., 1) forces worker processes to restart after a number of tasks, ensuring CUDA contexts are released.
• Impact / Considerations: Allows per-task process respawn to reclaim GPU memory.
• Related issue: Issue: CUDA workers do not release resources after each evaluate_program when using spawn start method #330
• Addition: Simplified Config initialization and dict conversion to reduce boilerplate and improve readability.

[yuxuan-z19]

Python <= 3.11 does not support max_tasks_per_child, use spawn context instead (see docs)