Error with specializer when max_depth > 1

[breuleux]

When using the inferrer with max_depth > 1, meaning that we propagate values for more than one level, the specializer produces erroneous graphs. The problem occurs with test_pow10 and also with the following:

@specialize((int1,))
def test_fact(x):
    def fact(x, a):
        if x == 1:
            return a
        else:
            return fact(x - 1, a * x)

    return fact(x, 1)

The VM complains about a cycle in that case. When looking at the graph, some free variables also seem to point to inputs in the wrong copy, and the generation also seems a little non-deterministic. There is clearly a bug somewhere.

We will not try to fix this immediately, since max_depth=1 is probably the setting we want, at least for the time being, and no problems have been diagnosed yet in that situation.