Native x86 floating-point backend for statically translated SPU work

File: runtime/spu/spu_native_float_x86.h
Synthetic differential test: runtime/spu/tests/test_spu_native_float_x86.c

The backend implements FA/FS/FM/FMA/FMS/FNMS with SIMD operations. It retains
the SPU integer reference for extended-range inputs and ambiguous overflow.
Zero results are normalized to positive zero. It does not enable fast-math,
replace SPU extended range with IEEE NaN/Inf behavior, or fuse separate guest
operations. It is independent of title addresses, assets and engine timing.

Caller contract
Select this backend only on a CPU and OS with AVX2 and FMA support. Baseline
code and feature detection must remain in separately compiled translation
units without those instruction-set flags. Selection occurs at image entry.

Save MXCSR before entering translated work. Set round-toward-zero, DAZ, FTZ,
and masked floating exceptions. Restore the saved value after the complete
work unit, including guest halt/stop handling. Do not expose this scope to
unrelated PPU, audio, renderer or host code. Bleach's outer native-job wrapper
owns this scope; the helpers deliberately do not change MXCSR per operation.

DAZ already supplies zero for subnormal inputs. Repeating input normalization
before each scalar/vector arithmetic instruction is redundant under this
contract. The sign of an input zero cannot affect a nonzero finite arithmetic
result. Output normalization is still required for positive-zero SPU results,
and all exceptional inputs/results still follow the exact integer reference.

Build the selected units with AVX2/FMA, floating-point contraction disabled
except for explicit fused intrinsics, and dynamic rounding respected. Tested
locally with Clang -mavx2 -mfma -ffp-contract=off -frounding-math -O2.

Validation boundary
The differential suite checks60,526,848 operation lanes against the integer
reference, covering signs, zero/subnormals, normal/extreme exponents, overflow,
cancellation and pseudorandom inputs. These tests passed on the i5-14600KF.
They are not a hardware PS3 oracle or a cross-platform performance claim.
Whole-game speed must be measured separately without this test or a profiler.
The title compatibility header retains existing generated function names.
