osp_it_fetch() allocates an array of 256 folios and 256 single page
folios for every OBD_IDX_READ, and osp_it_next_page() frees them as
soon as that bulk has been consumed, only for the next fetch of the
same iteration to allocate them again.

The split leaks as well: the fetch assigns it->ooi_folios
unconditionally, while the free is only reached once a bulk has been
consumed, which a failed fetch never is. As a failing
dt_index_operations::next is retried for as long as it fails unless
LPF_FAILOUT was asked for, every retry after such a fetch loses 1MB.

Keep the array for the life of the iterator and let osp_it_fini()
release it, which takes the allocation out of the fast path and
leaves no way to lose it. It is published once it is complete only,
so that a retry never reuses a half filled one. The old content of
a reused folio does not matter: only ooi_valid_npages of them are
read and every lu_page is validated by its magic.

Ask for the folios with GFP_NOFS, as the pages were asked for before
the switch to the folio API. These index objects are read with the
locks of the caller held, where a reclaim which re-enters this
filesystem can deadlock.

Fixes: 2b80459bf164 ("LU-17916 osp: switch to struct folio")