b=1896:

b=1896:

Two fixes, both required for this bug:
 - destroy the import when the export is freed, not when it's disconnected
   (which might be due to recovery, and might be racing with ASTs to access
   the same import data).  (Robert's patch plus a comment fix and an
   EXPORT_SYMBOL.)

 - take a reference in ptlrpc_fail_export, because obd_disconnect always nets
   us -2 (all other callers have another reference, either from a request or
   a local-connect artifact), and we only "own" one -- the one in the hash
   table, which represents the network connection.

Also: macroize class_export_put and class_export_get for a major simplification
of export-refcount debugging.