Ok, I should've tried the idea that David came up with first - it does
in fact make the code quite a bit simpler, and indeed removes the need
for the previously introduced "**error_msg" argument that I hadn't
really liked anyway.

So, changes here are:
 * move setting the bad attr pointer/message into validate_nla()
 * remove the recursion patch since that's no longer needed
 * simply skip the generic bad attr pointer/message setting in
   case of nested nla_validate() failing since that could fail
   only due to validate_nla() failing inside, which already sets
   the extack information

johannes