cmd/compile: preserve guarded AMD64 FP intrinsics in LLVM

[zhouguangyuan0718]

On AMD64, Go SSA guards SSE4.1 rounding and FMA operations with runtime.x86HasSSE41 and runtime.x86HasFMA. Lowering their true branches to generic LLVM intrinsics loses that contract at the target boundary: baseline X86 codegen may emit external trunc/floor/ceil/roundeven/fma symbols.

Use the guarded X86 intrinsics added by goallc/llvm-project#48 for AMD64 while keeping the generic LLVM intrinsics on other architectures. Update LLVM codegen expectations accordingly.

Validation:

• rebuilt the complete Go toolchain against the matching local LLVM payload
• TestLLVM/codegen/llvm_math_intrinsics.go passed
• TestLLVM/codegen/llvm_trunc.go passed
• linux/amd64 default compile, llc, GoObj archive, and final link passed for test/llvm_float_intrinsics.go
• the same pipeline passed for test/fixedbugs/issue68322.go
• optimized IR retained the runtime CPU-feature branches; assembly used roundsd/vfmadd213sd only on guarded paths; linked binaries had no unresolved generic FP symbols

Depends on goallc/llvm-project#48. The two runtime cases remain graylisted until linux/amd64 CI can execute them. This PR is intentionally left unmerged for review.