mm/huge_memory: read the huge PMD entry once when splitting it

__split_huge_pmd_locked() re-reads *pmd several times to classify the same entry. The PMD page table lock is held throughout, so read it once into old_pmd and classify from that. No functional change