init: fix fork_prepare lock order for heap_main.arena_pages_lock

[dylan-conway]

_mi_process_fork_prepare takes heap_main.arena_pages_lock before per-heap h->arena_pages_lock. At runtime the order is the opposite: a non-main heap's mi_heap_ensure_arena_pages holds h->arena_pages_lock and calls mi_heap_zalloc_aligned(mi_heap_main(), ...) (arena.c:672→682→1498→778), which can recurse into heap_main->arena_pages_lock. AB-BA with a thread mid-first-touch when fork() is called.

Moved heap_main.arena_pages_lock to after the per-heap loop (the original placement comment was wrong — mi_heap_ensure_arena_pages short-circuits for heap_main at arena.c:675 and never allocates). _mi_process_fork_parent release order updated symmetrically.