Try to fix fma on windows

[N5N3]

This PR aims to fix runtime fma on windows by adding julia_fma(f) using Base.fma_emulated as the reference.
On other systems, julia_fma(f) call use fma(f) directly.
I use Base.fma_float to test julia_fma(f)'s precison. And it passed locally.
Since it only affects const propagation, I believe perfomance is not important.
We can remove this code if fma on windows get fixed in the future.
Close #43386.

[oscardssmith]

having written the Julia version of this, congrats on having the fortitude it takes to rewrite this! This looks like the best solution to windows being bad for now.

[N5N3]

@oscardssmith Linux32's runtime fma seems also broken.
Is it OK to skip the test on Linux32 in this PR? (I have no linux machine.)

[oscardssmith]

32 bit linux is passing those tests on master, so they should continue to be tested and pass.

[N5N3]

master doesn't test runtime fma (call Base.fma_float(a, b, c) directly).
I mean keep test fma and Base.fma_emulated, and skip Base.fma_float only.

[oscardssmith]

that should be OK.

[vtjnash]

LGTM. What is wrong with linux32?

[N5N3]

I'm not sure, runtime fma test passed locally using official 32bit 1.7.1 on wsl2.
But CI test kept failing. (So I skip it on Linux32).
The current error should be unrelated.

[N5N3]

All test passed after rebasing @vtjnash