cmd/compile: preserve file-backed data in LLVM IR

[zhouguangyuan0718]

Large embedded files are represented by an LSym with FileInfo and a nil P slice. The native Go object writer streams FileInfo.Name, but LLVM data lowering previously allocated Size bytes and copied only P, producing a zero initializer. This made encoding/json/jsontext FuzzCoder consume a zeroed embedded corpus.

Materialize file-backed data once in the LLVM data lowerer and share the cached bytes between type construction and initializer construction. Validate the FileInfo size against the LSym, reject the invalid inline/file combination, and check both short and long reads. Native object emission and Go semantics are unchanged.

Add focused compiler tests for raw IR contents and malformed lengths, plus a >1 KiB string and []byte go:embed test that exercises LLVM O2 GoObj emission, linking, execution, writable storage, and separation from the read-only string.

Validation on darwin/arm64:

• make.bash -llvm-version=23 -llvm-link=static (local development payload containing current Go ABI support)
• go test -tags="llvm23 staticllvm" cmd/compile/internal/ssa
• go test cmd/internal/testdir -run "^TestLLVM/llvm_file_backed_embed.go$" -count=1
• go test with llvmtoolexec default, gcflags selecting encoding/json/jsontext, -run "^FuzzCoder$" -count=1 (all 170 seeds passed)

Linux qualification remains on the revision-locked release selected by CI.