#732

Summary:

I have added a DFS recursive method , _handle_broken_semaphores, to the tf::Node class to remove deadlocks that happen because of exceptions thrown down the chain. This is done via calling it on all preceding nodes stored in _edges vector. The method is similar to tf::Semaphore::_release taking a reference to a SmallVector of Node * and storing the waiter nodes' pointers so they can be rescheduled removing the deadlock. There's also an unordered set to handle loops. I also added a new state, _is_broken, to the tf::Semaphore class which prevents acquiring a semaphore that has had an exception until it's reset.

Tests and benchmarks:

• for now i only added a tweaked version of the test case presented in the initial issue to the semaphore unittest.
• Measuring the time of the unittests as suggested in the guidelines showed no measurable performance hit, nor do i expect on as most of the expensive work is done when handling exceptions.

Notes:

This is my first time contributing and I would welcome any kind of feedback and I am open to adding more test/ documentation to my changes.