killing orphaned child processes

[bghira]

This pull request improves the reliability and correctness of process termination and child process cleanup in the ProcessKeeper, especially for cases where subprocesses may outlive their parent or ignore termination signals. It also updates and unskips several tests to reflect and validate the improved behavior.

Process termination and child process handling:

• Added a Linux /proc-based fallback method _collect_child_pids_from_proc to recursively collect all descendant PIDs when psutil is unavailable, ensuring that orphaned child processes are reliably found and terminated.
• Refactored _force_kill_process_tree to collect all child PIDs before sending any signals, preventing orphaned GPU-hogging processes if the parent exits first. The method now uses either psutil or the new /proc fallback to find children, and then sends signals to all collected PIDs.
• Updated the final cleanup step to use the collected list of child PIDs for sending SIGKILL, ensuring all descendants are forcefully terminated even if they have been reparented.

Test improvements and re-enabling:

• Unskipped and improved several tests in tests/test_process_keeper.py to validate process termination, crash detection, and graceful shutdown. Tests now use polling loops to robustly detect process state changes, and timeouts are adjusted for more accurate measurement of graceful termination. [1] [2] [3]