overflow bug in inv(::Complex)

[stevengj]

As discussed in #5112, inv(1e300+0im) gives 0.0 - 0.0im rather than 1e-300 - 0.0im. The problem is that the definition inv(z::Complex) = conj(z)/abs2(z) is not robust to overflows.

The right thing is probably:

• Use a specialized version of the robust division algorithm to compute 1.0/z for Complex{Float64}, or maybe specialize it for /(r::Float64, z::Complex{Float64}) if that is faster than r*inv(z) (I'm guessing not). (Just take the existing complex-division algorithm and simplify it for the case of a real numerator 1.0.)
• Use the naive algorithm with Float64 arithmetic for Complex{Float32} and Complex{Integer} types where there is no possibility of Float64 overflow.
• Continue to use the naive algorithm for BigFloat.

(Similarly for complex division in general.)

[stevengj]

cc: @ggggggggg