NDB_SHARE PREMATURELY DESTRUCTED BY BINLOG RESTART,DUE TO INCORRECT REF COUNTING

That patch enhanced ndbcluster_mark_share_dropped() such that
it destructed the share itself, instead of calling
ndbcluster_real_free_share() when the use_count reached '0'.

However we failed to call ndb_index_stat_free() before
ndbcluster_mark_share_dropped() called NDB_SHARE::destroy().

Instead of introducing that call into ndbcluster_mark_share_dropped(),
this patch does some restructuring to achive the same - with less
duplicated code:

- ndbcluster_mark_share_dropped() is enhanced to free the
  index_stat when the share is dropped.

- ndbcluster_real_free_share() is enhanced to require the
  share to be in the dropped state before it is called.

- An assert is added to NDB_SHARE::destroy() checking that
  the index_stat has been freed prior to call.