chore: resolve builtin/foreign attribute names to a shared enum

[asterite]

Stacked on #13520 — the first three commits are that PR; review only the last commit here.

Problem

The name in a #[builtin(name)]/#[foreign(name)] attribute was carried as a String through the whole pipeline and re-matched textually at every consumer: the comptime interpreter's ~150-arm dispatch, the monomorphizer's HandledOpcode, Definition::Builtin/LowLevel, Intrinsic::lookup in SSA generation, and the ownership pass's clone-elision list (added in #13520). Nothing tied these tables together except one test, and an unknown name silently survived to SSA generation where it panicked.

Change

Builtin/foreign names now resolve once to noirc_frontend::shared::Builtin — a single enum (the two attribute kinds share one namespace at consumption) covering all 148 stdlib #[builtin] names and the 12 #[foreign] names, grouped by consumer: SSA intrinsics, monomorphizer-evaluated, comptime-only. The ACIR black box functions are not repeated as variants: they nest as Builtin::BlackBox(BlackBoxFunc), so acvm's enum stays the single source of those names and a new black box function is a Builtin by construction. strum derives provide lookup/name/iter for the unit variants (the payload variant is #[strum(disabled)] and handled explicitly), with a round-trip test.

Attribute tokens still hold free-form strings, so malformed or unknown names parse normally; resolution happens at the semantic boundaries:

• the comptime interpreter's call_special (then call_builtin/call_foreign match enum variants),
• the monomorphizer when constructing Definition::Builtin(Builtin)/LowLevel(Builtin),
• SSA generation maps Builtin → Intrinsic via the new Intrinsic::from_builtin (Intrinsic::lookup(&str) remains only for the SSA text parser).

Oracle (#[oracle]) names stay strings deliberately: they are an open, user-extensible namespace.

Behavior changes

• Calling a builtin whose name the compiler doesn't implement is now a proper MonomorphizationError::UnknownBuiltin diagnostic (with test); previously it was an SSA-gen panic!.
• The five unresolved_type_* comptime dispatch arms and their helpers are deleted: stdlib no longer declares those builtins, and only stdlib may declare builtins, so they were unreachable.
• BlackBoxFunc::AND/XOR/RANGE are filtered out of Builtin::lookup: they are not callable functions (the compiler emits their opcodes from binary ops, casts and range checks), so names like #[foreign(and)] — never declared anywhere — now error like any unknown name.

Follow-on wins

• The clone-elision sync test from fix(ssa): decide clone elision in the ownership pass, keeping clones mutated by sibling arguments #13520 now iterates Builtin::iter() exhaustively instead of a hand-maintained name list, and builtin_supports_clone_elision matches variants.
• One test adapted: the proxies test used a fabricated #[builtin(foo)], which the new diagnostic rightly rejects; it now uses the real generic builtin black_box and tests the same instantiation behavior.

Validation

noirc_frontend + noirc_evaluator (4192 tests), the full nargo_cli --test execute suite (9350 tests), stdlib tests (431 × 9 configs), AST fuzzer smoke, clippy, and the enum round-trip/black-box coverage tests all pass.

🤖 Generated with Claude Code