make dot docstring match the implementation

[adienes]

as written it's incorrect since sum will promote the types but + (and dot) overflow

julia> x = collect(0xf0:0xff); y = copy(x);

julia> sum(dot(i, j) for (i, j) in zip(x, y))
0x00000000000004d8

julia> dot(x, y)
0xd8

closes JuliaLang/julia#15690

[stevengj]

Should we instead fix the code to match the docstring?

[adienes]

I suppose that depends on whether you think of dot more as a reducing operation or a "scalar" operation on scalars-that-are-actually-vectors, in that reducing operations typically widen types and arithmetic does not.

note that numpy also chooses to overflow on np.dot (so Julia and numpy are at the moment consistent)