The bug was based on wrong undo data in recovery file and not enough checking of bad data.

sql/sql_select.h:
  Added comment
storage/maria/ma_blockrec.c:
  - Removed wrong sanity checks (didn't work for UNDO records)
  - More sanity checks and DBUG_ASSERT
  - More DBUG_ENTER and DBUG_PRINT
  - Removed filler blocks in extent_to_bitmap_blocks() as it caused problems in write_block_record().
    This was the main cause of the bug.
    (This change can make records generated by UNDO slightly smaller than original record, which we have to fix
    by correcting row_pos.length before calling write_block_record())
  - Fixed some problems in write_block_record() while doing UNDO.
  - Store header_length without TRANSID_SIZE into recovery log (as UNDO entires doesn't have TRANSID_SIZE)
  - Mark table crashed if something goes wrong during UNDO
storage/maria/maria_def.h:
  Added header_length