Scope strict FP to the decompiled sources

[bassdr]

Follow-up to #6811: scope the no-FMA / no-fast-math flags to soh/src (decomp) instead of the whole build. IS_ZERO removal retained.

Build Artifacts

• soh-mac.zip
• soh-windows.zip
• soh-linux.zip

[bassdr]

I re-tested the same soft-lock described in #6809. It's still fixed. I also experimented enabling LTO and no issue so far. Might be a way to get better performances on arm64 without sacrificing original console precise maths. I did not mesure the perf gain yet though.

What do you think?

[serprex]

I don't think it's worth it. We often copy decomp code into hooks etc, it'd be a weird error to figure out. Maybe scoping it to C++ in libultraship would avoid trouble. Ideally audio processing etc has simd intrinsics or something anyway. C99 introduced fma function to math.h, so code can always apply optimization itself

original mac PR scoped it to soh/src, but we decided to apply it globally since audio perf wasn't negatively impacted

[bassdr]

Closing, will experiment adding intrinsic or simd per-function optimisation instead. Indeed, audio usually benefit from fast math without compromising precision, but it's not a general rule elsewhere in the codebase. #6668 's modern audio pipeline is one of these candidates.