Base: twiceprecision: optimize mul12 (#49568)
It is well known and obvious that the algorithm behind `Base.mul12`
(sometimes known as "Fast2Mult") doesn't require a "Fast2Sum" (known in
the Julia codebase as `Base.canonicalize2`) call at the end, so remove
it. This follows from the fact that IEEE-754 floating-point
multiplication is required to be well-rounded, so Fast2Sum can't change
the result.
Reference, for example, the beginning of
https://doi.org/10.1145/3121432 by Joldes, Muller, Popescu.
Furthermore, `Base.Math.two_mul` already exists, so use it as a kernel
function. This required adding a general fallback method for
`Base.Math.two_mul`, required, for example, for `BigFloat`.
Also removed the `iszero` check that is now obviously not necessary.
