Remove the subnormal term in the improved rounding model

The improved rounding model can be rewritten as
```
rnd_b(x) = p_2(x, b) * eps(b)
```
where `p_2(x, b) = p_2(x)` if `|x| = 0` or `|x|` is in the normal range and `p_2(x, b) = min_pos_float(b) / eps(b)` when `|x|` is subnormal (and not `0`).