runtime: Use `__builtin_signbit` for mingw signbit
On i686, mingw-w64's math.h defines an inline __signbit that returns 0 or
0x200, while the out-of-line libmingwex version returns 0 or 1. GCC 15 emits
calls to the out-of-line function but folds `!signbit(x)` assuming the inline
definition's result, so `!signbit(a) == !signbit(b)` always compares false vs
true. This breaks the sign comparison in the round-to-odd step of
`fma_narrow`, making the runtime Float32/Float16/BFloat16 `fma_float` round
in the wrong direction whenever the rounding error and the result have
opposite signs.
CI failure: https://buildkite.com/julialang/julia-pr/builds/3179#01a0f076-7498-49fa-95f0-cac9923f1db9
Assisted-by: Claude Code (Claude Opus 5.5)
