Faster dict iteration

[JukkaL]

Using PyDict_Next would likely be a faster way to iterate over dictionaries. Implementing this in the IR seems a little tricky, since the API call mutates multiple registers, whereas primitive ops are expected to only mutate a single register. Returning a tuple internally might be a reasonable way to support this.

For more information, see https://docs.python.org/3/c-api/dict.html#c.PyDict_Next.

[msullivan]

We now use the "return a tuple internally" strategy for dealing with exc_info tuples in the exception APIs and it works well, so that is probably the way forward for this.