Expand interpreter coverage: references, match, intrinsics, aggregate format

[BornPsych]

Stacked on top of the import-interpreter-crate PR. This teaches the interpreter the constructs it previously rejected as Unsupported, so they stop being silently excluded from the cross-field differential. Every addition is field-independent and verified against Noir's ACVM executor: 69 bn254 tests, 51 goldilocks tests, clippy and fmt clean, and the ~505-program execution_success corpus survey at 0 MISMATCH / 0 FALSE-REJECTION (395 agree, up from 383).

The six commits are sequenced so the feature set reads as a progression:

• Back tuple fields with shared cells and add a reference value — the Value foundation (cell-backed tuples plus a Ref value) that all aliasing builds on.
• Interpret references and struct-field aliasing — &mut parameters, &mut s.field, and nested/multi-level references, including the &mut a.b.c-through-a-reference case.
• Interpret match and enum patterns — the match arm over the (tag, payload…) enum layout, with integer, boolean, and field scrutinees.
• Dispatch pure and field-independent intrinsics — slice ops, string/byte conversions, and field decomposition, plus the hint builtins (black_box, as_witness, assert_constant) and the range-check and bit-spread builtins.
• Interpret aggregate equality and format-string messages — ==/!= on arrays, tuples, and structs, and assertion f"…" message rendering.
• Cover closures, wire fixtures into the oracle smoke test, and document the roadmap — closures already work through the existing tuple + call machinery, so this locks that in and records the remaining coverage gaps.