Fixes the following assertion thrown by TBB:

	Attempt to terminate non-local scheduler instance

This took a significant amount of time to track down due to a number of
factors. If your looking for a quick fix; set your
tbb::task_scheduler_init to use thread local storage.

This is caused by using Python multi-processing (in our case for the
system tests). The process will spawn correctly (Windows), however
Python will helpfully create seperate thread(s). I assume to observe. In
this case the task_scheduler_init handle would then belong to one of
these threads, as the act of importing our Python module would init the
framework and TBB implicitly.

When the tests run and TBBs scheduling is required it will "sign_on" in
tbb's governor.cpp to handle scheduling. This works because the
scheduler_init handle
is process wide. However internally they then use thread local storage
(TLS) for tracking the scheduler.

As the main process goes to destruct it assumes that we must have
a scheduler instance, since at process level we have done the init.

But we cannot see any schedulers since they belong to a different
thread. We have been getting away with this as the TLS returns NULL
which effectively nops the destruction. But the assert correctly points
out were getting this wrong.

Marking each handle to the scheduler as thread local effectively ties
the destruction of the thread to its init status. For most threads this
effectively means we create a unique_ptr which isn't used. But for our
main threads and weird edge case worker threads it keeps the thread
state consistent.

We've used a unique_ptr since we have an API which allows us to change
the number of threads TBB can spin up. However, anyone else facing this
issue would probably be fine with a thread local object allocated on the
stack.