The existing parameters are currently pretty good and there's not much
point in changing them.

This switches between a Taylor-esque approximation based on x + x/2,
and the "obvious" exp(x) - 1 based on the size of x, with the exact
polynomial and the size threshold chosen by `optimiser`.

Fixes #10.

This switches between the linear approximation log_B(x) ~ x/ln(B) for
small x and the "obvious" log_B(1 + x), with the size threshold chosen
by `optimizer`. There was not much value in optimizing the
multiplicative constant (especially true for `ln_1p` if that is every
implemented: ln(x) ~ x avoids a multiplication), or using a
higher-degree polynomial, because the equation above is as accurate as
the `log2` approximation.

Fixes #11.