Fix a potential deadlock because of Eigen thread pool

Fix a potential deadlock when we do not notify Eigen thread pool of a new task, and forget to steal it when exiting the spin loop.

Imagine we submit 16 tasks that must run together to make forward progress (example is a rendezvous in XLA), and if we forget to wake up a thread for one of the tasks, then we get deadlock as 15 tasks will be blocked by the one task that will sit forever in the queue.