floating point `log` methods do not round correctly

[tspiteri]

The methods f32::log and f64::log do not round correctly. As a trivial example, this code

fn main() {
    println!("{}, {}", 13f32.exp2().log2(), 13f32.exp2().log(2.0));
    println!("{}, {}", 29f64.exp2().log2(), 29f64.exp2().log(2.0));
}

outputs

13, 12.999999
29, 29.000000000000004

[tspiteri]

There is no LLVM intrinsic for logarithms to an arbitrary base, probably because IEEE754 does not include that in its list of recommended correctly rounded functions. So maybe it's not a good idea to try to implement this with correct rounding. However, considering that the other transcendental functions are correctly rounded, the documentation should clearly state that the log methods are not, maybe something like Python's documentation. I would write something like:

Returns the logarithm of the number with respect to an arbitrary base, calculated as self.ln() / base.ln().

If the base is 2 or 10, the log2 and log10 methods can yield more accurate results.