Fix loop-carried results, Array.shift GEP stride, and mismatched-catch rethrow

[ASDAlexander77]

Summary

• ForOpLowering dropped loop-carried results for infinite for(;;) loops — the NoConditionOp branch never assigned the replacement ValueRange, so replaceOp ran with an empty list whenever the loop carried results.
• ArrayShiftOpLowering computed the post-shift memmove source offset using a hardcoded i32 GEP stride instead of the array's real element type, corrupting Array<number>.shift() (and any element type wider than i32).
• TryOpLowering's typed-catch RTTI mismatch (non-Windows/Itanium path) silently fell through to normal post-try control flow instead of rethrowing — the code carried its own // TODO: when catch not matching - should go into result (rethrow). Now rethrows via the existing NullOp+ThrowOp idiom, routing to the parent try's landing pad when one exists.

Found via a static audit of LowerToAffineLoops.cpp / LowerToLLVM.cpp (neither file had been touched since PR #198). Full findings, including 7 additional lower-confidence issues tracked for follow-up, are written up in docs/LowerToAffineLoops-LowerToLLVM-review.md.

Test plan

• Added 00for_infinite_result.ts (AOT + JIT) — infinite for(;;) with a loop-carried sum, asserts correct result.
• Added 00array_shift.ts (AOT + JIT) — Array<number>.shift() asserts correct return value and remaining array contents.
• Added 00try_catch_mismatch_rethrow.ts (AOT only — see note below) — nested try/catch where a mismatched typed catch must rethrow past it to an outer handler.
• Full ctest suite green: 341/341 JIT, 345/345 AOT.
• Verified the rethrow fix's IR directly via --emit=mlir-affine --mtriple=x86_64-pc-linux-gnu (correct ThrowCall/ThrowUnwind routing, including re-entry into an outer try's landing pad for the nested case), since this Windows JIT host only exercises the SEH path natively.

Note: the JIT variant of the mismatch-rethrow test is commented out in CMakeLists.txt — it hits a separate, pre-existing JIT limitation (Windows SEH landing-pad plumbing exits with code 3, no output) that reproduces identically on unmodified main, unrelated to this fix (which only touches the non-Windows/Itanium cmpValue path). The AOT variant passes and exercises the fix correctly.

🤖 Generated with Claude Code