Optimize i128 layout

FixedPoint based on i128 promotes to i256 during calculations. It significantly reduces performance, e.g.:
```
F64p9/rmul (~1e4, Nearest)               1.00      1.9±0.00ns
F128p18/rmul (~1e4, Nearest)             1.00    134.1±0.79ns
```

* Look for a more efficient implementation of i256
* Try to avoid i256 arithmetic when possible