A blob is stored in ndb in multiple parts. For a blob read, one
read operation is executed per blob part. If a blob part is not
found, the read fails with a 'row not found' error. This error
indicates a corrupted blob, since a blob should never have a missing
part. However, the 'row not found' error is reported as the
overall result of the blob read. The handler ignores such errors
and reports no error and zero rows.

This is fixed by adding a check for the case where a blob part is
not found. In this case, the 'row not found' error is overwritten
to set a 'corrupted blob' error. The 'corrupted blob' error is read
by the handler and causes the select to fail as expected. An
additional fix is added for the getValue() API to prevent the
'corrupted blob' error from being overwritten.

Test cases are added in mtr to test various types of SELECTs. An
autotest test case is added to count round-trips and check
corrupted blob reads using both blob read APIs - readData() and
getValue().

Change-Id: I456b953cc054eba78c8a46849a2a169577d56aeb