Perf: release swept GC cells through slab batches

Problem

The current object_churn profile is dominated by young-GC sweep and GcCellBacking.freeFn. Each dead object returns through the generic allocator path, reclassifies the same 1024-byte owned slab bucket, rechecks chunk ownership, and takes a bucket lock in parallel mode.

Parent: #62 Collector dependency: zig-utils/zig-gc#12

Proposed slice

• Implement zig-gc optional batched dead-cell release hooks in the Binding.
• Let GcCellBacking hold its bucket synchronization for the release phase.
• Release each exact owned cell directly into the appropriate free list without allocator-vtable dispatch.
• Keep finalizers ahead of the batch so their side-storage frees cannot recurse into held cell-bucket locks.
• Preserve the generic free path for non-cell allocations and embedders without the hook.

Acceptance criteria

• Focused GcCellBacking tests prove exact per-bucket counts, slot reuse, and delegated frees remain correct.
• Full zig-gc and zig-js suites pass with zero leaks; benchmark harness passes.
• Order-balanced fresh-process object_churn A/B preserves checksum 1030552128 and shows a repeatable improvement.
• Both small conventional commits cross-reference their issues and are pushed as Chris without coauthors.

Rejected and fully reverted together with zig-utils/zig-gc#12. The batched design regressed median-of-process-medians by 0.30% and won 2/7 pairs; the refined immediate direct-release design regressed by 0.99% and won 3/7 pairs. All checksums were exact and focused correctness gates passed. No source commit was created. This rules out allocator dispatch and batching alone; #62 remains open for a structural reduction in object initialization or sweep work.