Add monomorphized N-way zipper merge with frontier-based dispatch and specialization

[marcin-rzeznicki]

This PR introduces a generic N-way merge algorithm, extending the existing 2-way and 3-way implementations to arbitrary arity (well, <= 64) while preserving performance characteristics via specialization and careful traversal design.

N-way zipper merge (zipper_merge_n_mono)

• Fully monomorphized over const N and the zipper type (SomeMutRefZ)
• Uses a bitmask (active) to track participating zippers
• Stackless traversal using zipper movements and explicit depth tracking

Frontier-based traversal

• At each step:
• • Compute minimal byte across active zippers
• • Build a frontier (subset sharing that byte)
• • Dispatch based on frontier size:
• • • |frontier| = 1 → graft (no descent)
• • • 1 < |frontier| < N → subset recursion / specialization
• • • |frontier| = N → fast-path full descent (no recursion)

Specialized fast paths

• Dispatch to optimized implementations for:
• • 2-way → zipper_merge
• • 3-way → zipper_merge3
• • 4-way → zipper_merge4
• Larger frontiers recurse into the same function with a reduced mask

Introduce macro-based “variadic” frontends and tuple-driven helpers for composing multiple zippers in a single call.

• New macros:
• • zipper_meet_n!
• • zipper_join_n!
• • zipper_subtract_n!
• New helpers:
• • meet_n, join_n, subtract_n

Syntax

Tuple-based argument passing (&mut r1, &mut r2, ..., &rN).meet_n(&mut out) with macro-generated impls up to 64 (current limit: 32)

or

zipper_join_n!(r1, r2, r3, ... => w); instead