Problem
=======

For certain tables with big rows and with big primary key, taking backup
(snapshotend) while doing insert into table may hit some ndbrequire in data
nodes.

For snapshotstart backup concurrent deletes may hit the same ndbrequire.

The following ndbrequire have been seen failing in test:
Error object: BACKUP (Line: 10218) 0x00000002 Check sz <= buf.getMaxWrite()
failed
Error object: BACKUP (Line: 10448) 0x00000002 Check entryLength <=
trigPtr.p->operation->dataBuffer.getMaxWrite() failed

In Backup::execDEFINE_BACKUP_REQ() the data size allowed in a log entry is
set to 4096 words:

  const Uint32 maxInsert[] = {
    MAX_WORDS_META_FILE,
    4096,    // 16k
    BACKUP_MIN_BUFF_WORDS
  };

For backup redo log the log event for insert contains both primary key and
complete row (including primary key again) with attribute words.
In worst case that adds up to 5067 words, MAX_ATTRIBUTES_IN_INDEX(32) +
MAX_KEY_SIZE_IN_WORDS(1023) + MAX_ATTRIBUTES_IN_TABLE(512) +
MAX_TUPLE_SIZE_IN_WORDS(3500).

On top of that the log entry it self also use 4 words in header and possibly
a gci word and for undo log also a trailing length word, adding up to 5073
words.

Solution
========

Define a new constant MAX_BACKUP_FILE_LOG_ENTRY_SIZE and use that in
maxInsert[1] to set max words in an insert log entry.

Reviewed-by: Frazer Clement <frazer.clement@oracle.com>