Fix bug in finding arity for verify attributes

[bakaq]

This fixes a bug in the current attributed variables implementation that creates a lot of strange bugs that seem like miscompilations.

Closes #2706, closes #2809, closes #2632.

[adri326]

I don't have much experience with the logic behind verify_attributes, but I couldn't find any issue with this fix.

I wonder how hard it would be to store somewhere else the number of temporary variables, so that the WAM instructions don't need to be scanned for every attributed variable unification.

---

I do believe that the current method of modifying self.code to eventually call verify_attr_interrupt is iffy. It feels flimsy and it will make JIT impossible in its current form.

[bakaq]

I wonder how hard it would be to store somewhere else the number of temporary variables, so that the WAM instructions don't need to be scanned for every attributed variable unification.

I do believe that the current method of modifying self.code to eventually call verify_attr_interrupt is iffy. It feels flimsy and it will make JIT impossible in its current form.

I think ideally this should all be done in the compilation step, although maybe the verify attributes interrupt really needs to be this dynamic.

[bakaq]

I made the code more functional, put some asserts and comments, and now scan the entire current predicate for use of temporary registers. I'm not sure if that is needed, but I think it's better to be conservative with this, because saving more registers than necessary will never be wrong, just inefficient.

If someone who understands how this works can assure me that we only need to scan from machine_st.p to the end of the predicate this is very easy to change.