mul_add documentation is inaccurate

[vks]

According to the documentation of mul_add:

This produces a more accurate result with better performance than a separate multiplication operation followed by an add.

This seems not to be true in general. On my machine, a.mul_add(b, c) is slower than a * b + c when compiling without target-cpu=native.

[vks]

cc @pitdicker

[hanna-kruppe]

cc #44805, dedicated FMA instructions are not universally available and we can't very well replace a.mul_add(b, c) with a * b + c when no such instruction exists (it rounds differently and also has different behavior when NaNs or infinities are involved).