fix(hir): class destructuring (dstr) + default-parameter parity (test262)

[proggeramlug]

Summary

Brings language/{statements,expressions}/class/dstr from 512 → 664 passing (41.6% → 53.9%) with zero regressions, and lifts the wider built-ins+language sweep by ~6–9% per shard (0 regressions across 4 sampled shards 0,3,6,9 of 12). These are method / constructor / accessor parameter-destructuring tests across plain, generator, async-generator, static, and private class methods.

Root causes fixed

1. Array binding patterns now use the iterator protocol. lower_pattern_binding lowered let [a,b] = x (and every method/param destructuring) to raw index reads (x[0], x[1]) — never invoking Symbol.iterator, never closing the iterator, mishandling holes/rest, and unable to destructure non-array iterables. Rewrote the Pat::Array arm to emit GetIterator → IteratorStep/IteratorValue per element, drain a rest element via a new js_iterator_rest_to_array, advance the iterator on elisions, and IteratorClose on both normal completion (when not exhausted) and any abrupt completion. Destructuring defaults now use strict === undefined (a genuine NaN element no longer triggers the default).
2. Object binding patterns enforce RequireObjectCoercible (new js_require_object_coercible) — TypeError for a null/undefined source even for an empty pattern {}.
3. Static-method calls pad omitted args with undefined. Expr::StaticMethodCall (non-rest path) forwarded only supplied args, so a static method called with fewer args read uninitialized parameter slots — static f(a=1) returned 0, static m([x]=[]) threw.
4. Private methods emit their default-parameter prologue — lower_private_method computed param.default but never called build_default_param_stmts.
5. Method-as-value calls pad omitted args with undefined — call_vtable_method padded a bare IEEE NaN; reached without call-site padding when a method is invoked as a value (const f = obj.m; f(), or a get m(){ return this.#m } accessor exposing a private method).

Validation

• Cluster (class/dstr): 0 regressed, 152 fixed vs main baseline.
• Wider sweep (built-ins+language, shards 0/3/6/9 of 12): 0 regressed, +20/+20/+19/+26 fixed.
• Built + tested on a Linux box against the pinned test262 corpus.

Remaining class/dstr failures are concentrated in *-init-fn-name-* (NamedEvaluation of anonymous function/arrow/class defaults) and generator throw-timing (eager param destructuring) — distinct features left for follow-ups.