x87: Fix FPREM/FPREM1 semantics in dynarec and interpreter

[yzewei]

FPREM is now calculated using fmod, and the lower-order quotient is calculated using trunc(x/y) to write back the condition code bits (C0/C1/C2/C3).

FPREM1 is now calculated using remquo, ensuring that the remainder matches the lower 3 condition code bits of the quotient.

Special inputs such as NaN/Inf/division by zero are handled separately, and abnormal behavior is aligned with the x87 expected flag.

[ptitSeb]

Can you check your changes with games llike Flatout / Flatout 2 ?
On what program/games did you check this PR?

[yzewei]

Can you check your changes with games llike Flatout / Flatout 2 ? On what program/games did you check this PR?

I discovered this while running NumPy's official test cases test_float_remainder_errors and test_float_remainder_overflow. I don't have the game environment running here.

[yzewei]

Can you check your changes with games llike Flatout / Flatout 2 ? On what program/games did you check this PR?

However, it is certain that this problem can be reproduced.

[ptitSeb]

Ok, I'll check later, and roll back the change if I notice regressions (this opcode is sensitive)

[yzewei]

Ok, I'll check later, and roll back the change if I notice regressions (this opcode is sensitive)

No problem, thank you!