Triggering an assert in a thread leads to calling siglongjmp to a
jmpbuf created from another thread (the main thread). That's undefined
behaviour and breaks all kinds of things.

Other issues with threads currently: Interleaved backtrace printing,
stack corruption because the longjmp succeeds in another thread and
continuing execution modifies the stack for the other threads that are
still running and triggering their asserts, no good automatic way of
propagating failures to main thread otherwise.

We probably need a whole framework for igt_thread á la igt_fork to
handle all that.

Signed-off-by: Petri Latvala <petri.latvala@intel.com>