Tests were failing because in TIME_from_longlong_datetime_packed() GCC8
at -O2 assumed that tmp is always positive and used mul and shr while it
used imul and sar at -O1 (where tests passed). GCC8 used multiplication
(by 0x4ec4ec4ec4ec4ec5) and shift to implement division by 13. It could
assume that tmp is always positive, because the function starts with
`if (tmp < 0) tmp= -tmp;`

But this assumption breaks if tmp=0x8000000000000000;
This is invalid value and TIME_from_longlong_datetime_packed() should
never see it, garbage in - garbage out.

It was getting this invalid value because mroonga tried to convert a
NULL key part to MYSQL_TIME. If the key part value is NULL, datetime2
value of it happens to be bzero-ed, which is invalid binary datetime2
value.

The correct behavior is not to try to interpret the key part value, if
it is marked as NULL. But this minimal fix only covers the datetime2 type.