Problem :
---------
In ha_innopart::rnd_next_in_part() and in other partition fetch calls we
are freeing the current blob heap for the partition before doing record
fetch. It looks like SQL is referring to this BLOB heap in case there
is no data to fetch and so freeing it early causes the issue in
correlated sub query execution. The issues is introduced from
[hash:eb16156bf16388153a27e776439dee38a18b2a4b] as part of the bug-fix
[bug-20582149 - INVALID READ OF SIZE 1 IN PROTOCOL::NET_STORE_DATA].

Solution :
----------
It is good to keep the part consistent between partitioned and
non-partitioned table. The original fix doesn't need to free the blob
heap before doing fetch.

1. clear_blob_heap_part() - Removed function and calls from next and
   prev for index and sequential scan.

2. For table scan and unordered scan use only one blob_heap.

3. ha_innopart::rnd_end_in_part() - Remove unconditional reset of
   blob_heap to NULL.

Reviewed-by: Mattias Jonsson <mattias.jonsson@oracle.com>

RB: 8215