build: use SSE math on 32-bit x86 also when the tarball name is given

[maleadt]

#43978 added -mfpmath=sse to JCFLAGS on 32-bit x86, so that Julia's C runtime does
not use 80-bit x87 math. It went inside the ifeq ($(JULIA_BINARYDIST_FILENAME),)
block that computes the binary-dist tarball name, so the flag is dropped whenever
that variable is already set. CI sets it: julia-buildkite's build_envs.sh exports
it. The i686 Linux and Windows CI builds and nightlies have therefore compiled
src/*.c with x87 math all along (make ARCH=i686 print-JCFLAGS includes
-mfpmath=sse, and JULIA_BINARYDIST_FILENAME=x make ARCH=i686 print-JCFLAGS does
not).

The x87 build now fails a test once Windows i686 is built with GCC 15.2 and
BinaryBuilder's MinGW headers (the planned CI toolchain, matching
CompilerSupportLibraries). Base.fma_float at runtime, which on Windows uses the
emulated julia_fma in runtime_intrinsics.c, returns NaN for
fma(floatmax(), 2.0, -floatmax()) and similar inputs (test/math.jl:1551-1567).
Those headers classify floats with an out-of-line __fpclassify, which sees the
rounded double, while the rest of the function computes in x87 extended precision.
The mingw-builds GCC 12 used so far inlined an fxam on the unrounded value
instead, which matched. With -mfpmath=sse the results are correct either way;
without it, a standalone copy of the function gives nan with
-D__CRT__NO_INLINE.

This moves the flag out of that block, so it always applies on 32-bit x86.

Assisted by: Opus 5.5

[maleadt]

Merging early, as this is Windows-only and mingw23-build and -test passed CI.