If the record is not found during rollback, attempt to insert it,
unless a duplicate key error would occur. If a duplicate key error
would occur, flag the index corrupted in the data dictionary cache.

The condition is perfectly normal during online secondary index
creation. We have seen it also during stress testing that does not
involve online index creation, but the root cause of that bug is
unknown.

row_undo_mod_del_unmark_sec_and_undo_update(): Suppress the log
printout about missing record, when the index was being created
online. Attempt to insert the missing record. Return DB_DUPLICATE_KEY
when this would result in inserting a duplicate.

row_undo_mod_sec_flag_corrupted(): New auxiliary function, for
flagging an index corrupted during rollback.

row_undo_mod_del_mark_sec(): Flag the index corrupted on
DB_DUPLICATE_KEY, but set err=DB_SUCCESS so that the rollback will
carry on as usual on the other indexes.

row_undo_mod_parse_undo_rec(): Remove a redundant parameter.

row_ins(), row_upd(): Add test instrumentation.

dict_set_corrupted(): Acquire the data dictionary latches when
needed. Add a context parameter, so that we can log where the
corruption was noticed.

rb:1405 approved by Jimmy Yang