mysql-test/t/loaddata.test:
  test for bug; without fix, running the test with --valgrind would show the leak
  and make the test fail.
sql/sql_load.cc:
  * In READ_INFO class, 'need_end_io_cache' is true as long as init_io_cache() was called,
  so if it's true, we need to call end_io_cache(), to free memory allocated
  by init_io_cache(). No matter the value of 'error'. In the bug's scenario,
  'error' was set to true in read_sep_field() because
  '1' (read from file) isn't suitable to load into a geometric column. Because of
  'error', end_io_cache() was not called.
  Note: end_io_cache() calls my_b_flush_io_cache(), which will do nothing wrong given
  that the file is opened for reads only; see the init_io_cache() call which uses
  only those read-only types:
  (get_it_from_net) ? READ_NET : (is_fifo ? READ_FIFO : READ_CACHE).
  IF the cache were rather used to write to the file, my_b_flush_io_cache() may
  write to it, and it may be questionable to write to the file
  if 'error' is true. But here there's no problem.
  * Now that 'need_end_io_cache' is checked even if 'error' is true, it needs
  to be initialized in all cases.
  * Bonus: move some variables to the initialization list.