Add sum methods for BigInts and BigFloats

[andrioni]

As due to a pointer issue mpfr_sum isn't wrap-able, I've decided to write a slightly more efficient method for BigFloats (and another one for BigInts with the same structure) reusing the summation variable.
In my tests with arrays with 10^6 numbers, it seems to be about 5-10x faster than the generic version.

[stevengj]

Couldn't you implement + and * of multiple arguments in the same way (to avoid the extra temporary vars)?

Note that you can use mpfr_sum once #2818 lands. (Actually, there is a hackish way to use it now, by using jl_value_ptr combined with & get C-compatible struct pointers. Probably better to just wait for 2818, though.)

[andrioni]

Just did with also $, | and & for good measure, as they are the other commutative and associative operators we have in BigInt.

[stevengj]

Instead of defining + etcetera for 2, 3, 4, and 5 arguments separately, can't you just define function +(xs::BigInt...) to handle an arbitrary number of arguments?

[andrioni]

But we can't typecheck then, as these new methods are defined when all arguments are BigInt. Well, I could write a generic varargs method, but having to typecheck then kind of makes it moot.

[andrioni]

(maybe a varargs method typechecking to see if it's of one of the possible conversions and then doing it keeping the accumulator variable mutable could be worth it, but it'd have to basically replicate all external constructors.)