feat(codegen): const-fold literal new Function bodies to native functions #1679

Part of #1677. Blocked by #1678 — do not start until it is merged to main and parity-green.

Problem

When a new Function body is a compile-time constant string, an AOT compiler should turn it into a real native function instead of refusing it. This is true ahead-of-time eval and it builds the string → HIR plumbing reused by Phase 3.

Scope

For call sites the Phase 0 classifier marks const-foldable:

• Resolve the body (and any param-name args) to constant strings via the existing const-prop pass.
• Parse the body with SWC as a function body.
• Lower it to HIR exactly as if the user had written the equivalent function literal.
• Emit native code; replace the new Function call with a reference to it.

Note new Function has no access to local scope (globals only), so there is no closure capture to model — the lowered function is a plain top-level function referencing globals.

Covers: new Function('a','b','return a + b'), the (0, eval)('this') → globalThis idiom, trivial generated accessors with literal bodies.

Acceptance criteria (the gate for Phase 2)

• test-files for: literal single-expression body, multi-arg param names, indirect-eval globalThis idiom.
• Byte-for-byte parity vs node --experimental-strip-types.
• Non-constant bodies still fall through to the Phase 0 refusal (no regression).
• Merged to main, lint clean, parity-green.

Out of scope

Library-generated (non-literal) body strings — those are Phase 3.