WL#2424: Optimize Item_func_numhybrid()

Currenty we have a switch and one extra call for all 'val_xxx' operations that 
uses the Item_func_numhybrid class.

The current code is:

String *Item_func_numhybrid::val_str(String *str)
{
  DBUG_ASSERT(fixed == 1);
  switch (hybrid_type) {
  case DECIMAL_RESULT:
  {
....
  }
  case INT_RESULT:
...
}

This should be changed so that we when we calculate 'hybrid_type' set up 
virtual functions for calling the needed val_string, val_int, val_decimal and 
val_real.

In some cases we can call the 'real' evaluation function directly like in case 
of (current code):

double Item_func_numhybrid::val_real()
{
  switch (hybrid_type) {
...
  case REAL_RESULT:
    return real_op();

This will make operations like 1+1 in integer contest or 1.0+1.0 in real 
context as fast as possible.