fpu: canonicalize NaN results of fadd/fmul/fdiv and fused ops

[carlosqwqqwq]

Summary

The FP writeback for fadd, fmul, fdiv and the fused fmadd/fmsub/fnmsub/fnmadd family stores the raw host result, which keeps non-canonical NaN payloads. The fsub/fmin/fmax paths already canonicalize through the write helpers; the RISC-V specification requires the canonical NaN bit pattern for arithmetic results. Route the remaining arithmetic writebacks through the same canonicalizing writers.

Signaling-NaN NV flagging is unchanged.

Validation

• qNaN-payload, inf*0+1 and 0/0 probes match native RISC-V hardware and QEMU after the fix, with the JIT enabled and disabled.
• Every arithmetic result is the canonical NaN and fflags match the references.

Fixes #254

[LekKit]

@SolAstrius please confirm via your out-of-tree test suite and review if possible, thank you

[carlosqwqqwq]

I reran the fix at head 35fdcb4 with an out-of-tree RV64 probe, built independently from the PR source tree, under the Docker x86_64 lab. The probe output was:

fadd=0xffffffff7fc00000
fsub=0xffffffff7fc00000
fmul=0xffffffff7fc00000
fdiv=0xffffffff7fc00000

Thus the affected fadd/fmul/fdiv writebacks now match the canonical NaN reference. fsqrt is intentionally outside this PR (tracked separately as D-006). The PR head also passes the full GitHub Actions build/CodeQL/analysis checks. @SolAstrius, please confirm with the out-of-tree suite when available.