e.g. Float -> Double

We do not use the python convention for floordiv (always rounding towards
negative infinity), but instead round towards zero. The SQL backends and C does it
like that. Related to that, the sign of the result of % is determined by the first
operand, not the second (again, this is SQL / C convention, in python it is determined
by the second operand). A reason for using the C convention is that

                       -(a // b) == -a // b == a // -b

which is not the case for python.

the ones with extreme values are skipped for postgres

we do not allow t >> select() >> mutate(x=C.a). Columns available via C are
exactly the columns that are written on export().

SQL Server 2022 can do absolutely no nan / inf, so it's rather a minority of
backends that can do this. We thus declare this as implementation-dependent.

When the backend does not use i64 / f64 by default, cast to it. In the tests, we
now require to get the exact same data type on export to polars for all backends.