Float128 fma() subtly corrupts state on Windows

On Windows (but not Linux or OSX) calls to `fma()` somehow break the overflow handling.

```julia
julia> h, x = floatmax(Float128), Float128(1)
julia> h + h
inf
julia> fma(x,x,x);
julia> h + h
1.18973149535723176508575932662800701e+4932
```

I don't see anything wrong with our wrappers; it may be a libquadmath and/or libgcc issue.

Start by reproducing the Float128 overflow sequence on Windows and compare it with Linux or OSX. Inspect the Julia wrappers around fma() and investigate the suspected libquadmath or libgcc interaction. Done means calling fma() no longer changes subsequent overflow handling, with the comparison behavior verified across platforms.

Written by the indexing model from the issue text.