This error happens when we try to read the first record from the
database and a lock timeout occurs. This lock timeout error message
does not get correctly propagated. The assert happens when the reply
is going to be sent to the client and there is no error code set in
the diagnostic area. The reason this happens when MRR is used is that
when using MRR more work is done during initialization of MRR. One of
the things the MRR code does is to try to fill the row id buffer by
reading from the database index. This result in a call to
ha_innobase::multi_range_read() which returns the lock timeout error
code (error 146). This error code is not propagated up to the server code.

When MRR is not in use there will be no calls to the database during
initialization. The lock timeout error will only occur when the first call
to read a record is done. With this code path the error code is
correctly propagated to the server code and written to the diagnostic
area before the reply message get sent to the client.

This patch fixes this problem by the following two changes:

1. Fix DsMrr_impl::dsmrr_init() so that it returns the same error code
   as reported by the storage engine instead of just returning one when
   an error occurs.

2. Extend join_init_read_record() (in sql_select.cc) so that it
   handles errors that occurs during initialization (in this case from
   QUICK_RANGE_SELECT::reset()) and ensures the error code gets written
   to the diagnostic area.