Version for Feature Preview. 
This patch contains the following: 

1. Upgrade of the JOIN_CACHE structure to a class.

2. Optimization of the function that writes into the cache buffer:
    - null fields are not copied into the cache anymore
    - varchar values are packed with their lengths prepended,
    - if the first match is needed all records in the cache are,
      prepended with their length that allows to skip them easily.

3. The class JOIN_CACHE_BNL derived from the class JOIN_CACHE 
   that implements the Block-based Nested Loops Join algorithm.
   This algorithm was implemented earlier but unlike the existing
   implementation of the Nested Loops Join algorithm it did not
   support outer joins and first match joins. This implementation
   supports all such joins including nested outer joins.

4. The class JOIN_CACHE_BKA derived also from the class JOIN_CACHE.
   The class implements the Batched Key Access Join algorithm.
   The implementation uses the existing implementations of the
   Multi-Range Read Interface for MyISAM/InnoDb and for NDB Cluster.

Two varieties of join caches are supported: plain caches and linked
caches. Records in a linked cache can refer to the records in other
caches. So if two partially joined records were generated from the same
initial records they just refer to their common beginning. It allows to
save space in join buffers.

Although this is already the result of refactoring of the first prototype
implementation the code is still rather fragile. Hopefully the comments that
were added in abundance will help you to navigate through it. Any improvements
are welcome.  

Original revid:sp1r-igor@olga.mysql.com-20080401061749-11032