Optimize Rows_log_event Reporting of Process Info

Reporting on the status of Rows_log_event execution (as seen by process_list) is inefficient.

1. thd_proc_info(thd, message) is called within do_exec_row(), without the possibility of the message actually changing between rows; however, the message is re-built entirely using my_snprintf for each row. my_snprintf is slow and doesn't need to be called per-row. This could instead be done in do_apply_event() around the loop which iterates rows, thereby only needing to call my_snprintf() once.
2. When WSREP_PROC_INFO is defined, my_snprintf() is called again to re-write the same message into thd->wsrep_info. This should also be moved to do_apply_event.
3. In error checking, thd->is_error() should always be set when write_row() (or whatever equivalent for the Rows_log_event implementation) is called. This can be switched to a debug assertion DBUG_ASSERT(!error || thd->is_error) where we can just return the returned error.

The above logic should be applicable for all Rows_log_event implementations:: Write, Update, and Delete.

Note these points come from monty, I'm just filing the JIRA.

Depending on the speedup seen from this, it may be beneficial to put into 10.6 (rather than a new version, as cleanups usually are).

Note the above suggestion says that my_snprintf is slow and doesn't need to be called per-row; but actually it shouldn't be called at all. Notice the following comment for the proc_info variable (which is assigned this string):

  /*

    ...

    This member is accessed and assigned without any synchronization.

    Therefore, it may point only to constant (statically

    allocated) strings, which memory won't go away over time.

  */

  const char *proc_info;

Dynamically creating this message for SHOW PROCESSLIST violates this principle, potentially leading to crashes. When moving the thd_proc_info call, we should revert to the original message displayed before MDEV-7409 (and consequently revert MDEV-29894 entirely).