This applies to int and native integer types.

This speeds up this micro-benchmark by up to 80% (it spends most of the time
unboxing integers):

# a is list[int]/list[i64]/...
for i in a:
    if i == 789:
        n += 1

The impact to compile time when self-compiling is below the noise floor. The generated
binary is about 0.1% larger. Since integer unboxing can be performance-critical,
this seems like a decent win.

Closes mypyc/mypyc#987. Work on mypyc/mypyc#757.