Correct C++ floor division for integers

[mstimberg]

This fixes #1495. As mentioned in the issue, the problem is that we use floor(1.0*x/y) to calculate floor division with Python semantics in C++. By default, we have unsafe math optimizations enabled, which leads to this not always being correct (e.g. 98/98 would be rounded down to 0). I think this is due to replacing the division by multiplication with an approximated reciprocal, or something along these lines. Switching off unsafe optimizations would be an option, but it would slow down code in other places.
In this PR, I instead create function specializations for integer arguments, which then use a rather complex calculation to get the value:

q = a / b;
r = a - q*b;
q -= ((r != 0) & ((r ^ b) < 0));  // result

This is the implementation used by Cython (cdivision: false, obviously), and it should cover all corner cases. I didn't benchmark this, but it is probably quite a bit slower than the previous solution. This could bring up #993 on the table (the issue is about Cython, but the same is now relevant for C++), but on the other hand all use cases I can think of for floor division and modulo operations are initializations and run_regularly operations, nothing that is time-critical. As a data point, all usages of // in our examples are for initializations like x = i // columns.