Fix ~10x slowdowns on NUMA nodes with fragmented memory: numa_zeros() no longer requests transparent huge pages with MADV_HUGEPAGE when the kernel would compact memory synchronously to satisfy the request (transparent_hugepage/defrag = always, madvise or defer+madvise). With enabled=always the kernel still uses free huge pages; on fragmented nodes it now falls back to local 4 KB pages instead of stalling in compaction and placing huge pages on remote NUMA nodes.