perf(codegen): share recursive struct Drop helpers

[sanohiro]

Summary

• emit one module-private iterative destructor per Move struct reached as a Drop-site root
• reuse that helper across direct, replacement, fixed-array, dynamic-array, and cross-function cleanup sites without introducing recursive helper calls
• add malformed-input, sibling aggregate-array, and 4,096-level stack-bound owners plus a local baseline/candidate benchmark

Request 19 evidence

• raw LLVM IR: 1,517,324 lines / 113.6 MB -> 109,992 lines / 5.96 MB
• cold three-unit release build: 471.074 s and >832,704 KiB observed RSS -> 13.555 s and 266,400 KiB peak RSS
• final release candidate: 13.27 s, 264,560 KiB, exact prompt-verifier PASS output
• small control: unchanged 1,240-byte object and 1 allocation / 1 free; cleanup execution 0.001 s before and after

Verification

• cargo build --release --workspace
• scripts/cargo.sh test -p align_codegen_llvm --lib
• scripts/cargo.sh test -p align_driver --test large_drop_codegen
• Request 19 fixture built and executed with the final release compiler
• code-tier scripts/pre-pr.sh (focused owner, bounded gate, Clippy)