To improve the current cost estimation functions for reading data from
a table or index, information about whether table or index data is in
memory or need to be read from disk is needed. The worklog extends the
handler with:

a) Data structures where storage engines can provide estimates for
   how much of data from tables and indexes that are stored in a main
   memory buffer.
b) An API that the optimizer cost model can use to access this information.
c) A default implementation/heuristic that can be used if the storage
   engine has not provided this information.

In addition, the ha_heap::info() function for the Memory engine is
extended to provide these estimates.

Per file comments:

sql/handler.cc:
  Default implementations for the functions that provide estimates
  about how much of tables and indexes are stored in a memory buffer.
sql/handler.h:
  Extend the handler class with:
  a) data structure for the storage engines to provide estimates for
     how much of a table that is currently stored in a memory buffer.
     This is done by extending the ha_statistics class with a new
     variable.
  b) an API that the cost calculation functions can use for getting
     information about how much of tables and indexes are in a
     memory buffer.
sql/key.h:
  Extend st_key with a new variable and functions that can be used by
  storage engines for providing an estimate of how much of the index data
  that is currently stored in a main memory buffer.
sql/table.cc:
  Initialize the st_key::in_memory_estimate to a default value.
sql/sql_tmp_table.cc:
  Initialize the st_key::in_memory_estimate to a default value.
storage/heap/ha_heap.cc:
  Extend ha_heap::info() to provide estimates for how much of
  tables and indexes that are currently stored in a main memory buffer.
unittest/gunit/key-t.cc:
  Added unit test for interface for setting and getting in-memory estimates.
unittest/gunit/handler-t.cc:
  Added unit test for handler::table_in_memory_estimate() and
  handler::index_in_memory_estimate().