Add benchmark to stress specialization by doing different specialization on different threads.

[markshannon]

Take this fib function:

def fib (n, one):
    if  n <= 1:
        return one
    return fib(n-1, one) + fib(n-2, one)

Call fib(20, 1) on one thread, and fib(20, 1.0) on another.
Maybe use something other than fibonacci, but you get the idea.

With a GIL the specialization will flip-flop, but will be mostly correctly specialized as since the tier 1 interpreter is adaptive.
This will be a challenge for NoGIL, however.
In theory the tier 2 optimizer should work with NoGIL, but this should tell us if it doesn't.