We've been using __GFP_NOFAIL for allocating struct bch_folio, our
private per-folio state.

However, that struct is variable size - it holds state for each sector
in the folio, and folios can be quite large now, which means it's
possible for bch_folio to be larger than PAGE_SIZE now.

__GFP_NOFAIL allocations are undesirable in normal circumstances, but
particularly so at >= PAGE_SIZE, and warnings are emitted for that.

So, this patch adds proper error paths and eliminates most uses of
__GFP_NOFAIL. Also, do some more cleanup of gfp flags w.r.t. btree node
locks: we can use GFP_KERNEL, but only if we're not holding btree locks,
and if we are holding btree locks we should be using GFP_NOWAIT.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>