cmd/compile: lower float truncation to LLVM

[zhouguangyuan0718]

Summary

• lower Go SSA OpTrunc with the semantics-matching llvm.trunc.f64 intrinsic
• add SSA-checked LLVM IR coverage and runtime special-value/boundary checks
• qualify fixedbugs/issue68322.go for GoObj compile, link, and execution on darwin/arm64
• keep linux/amd64 GOAMD64=v1 fail-closed because its separate HasCPUFeature lowering remains unsupported; GOAMD64=v2 IR compilation succeeds

Validation

Revalidated after rebasing onto go1.27.master@5c34a612d8d5:

• go test -tags=llvm23,dynamicllvm cmd/compile/internal/ssa
• targeted Test/LLVM codegen for llvm_trunc.go
• targeted Test/LLVM runtime for fixedbugs/issue68322.go and llvm_float_intrinsics.go
• rebuilt Go toolchain and statepoint plugin against merged LLVM cmd/llvmplugin: describe pointer allocas as memory roots #33
• gofmt and git diff --check

Earlier focused qualification also covered opt -passes=verify, llc -verify-machineinstrs, direct GoObj build/link/run, and native math.TestTrunc.