bch2_path_put_nokeep() is sketchy, and we should consider removing it:
it unconditionally frees btree_paths once their ref hits 0.

The assumption is that we only use it for paths that have never been
visible outside the btree core btree code; i.e. higher level code will
never be making assumptions about locking based on these paths.

However, there's subtle brokenness with this approach:

 - If we call bch2_path_put(), then bch2_path_put_nokeep(),
   bch2_path_put() may free the first path on the assumption that we we
   have another path keeping a node locked - but then
   bch2_path_put_nokeep() just unconditionally frees it.

The same bug may arise if we're calling bch2_path_put() and
bch2_path_put_nokeep() on the same (refcounted) path, or two adjacent
paths that point to the same btree node.

This patch hacks around one of these bugs by calling
bch2_path_put_nokeep() first in bch2_trans_iter_exit.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>