fixed_point compare uses too great a range

In
```c++
auto a = fixed_point<int8, 32>(0x100000000LL);
auto b = fixed_point<int8, 0>(1);
ASSERT_GT(a, b);
```
expression overflows. But only 16 (maybe 15) bits of range are required to scale operands such that comparison can be performed. See [issue 123](https://github.com/johnmcfarlane/fixed_point/issues/123) for a discussion.