Profiling SequenceManagerPerfTest suggests that avoiding locks is a
useful optimisation.  We currently take |immediate_incoming_queue_lock_|
every time we post a task, even from the main thread.

Ideally we'd have a separate main thread only queue, and make 
TaskQueueImpl::ReloadImmediateWorkQueueIfEmpty reconcile the two task
sources (we should merge keyed on enqueue order).  Alternatively we could
do something clever in WorkQueue.