Fix cost calculation for get_best_group_min_max()

If the final range restrictions (SEL_ARG tree) over GROUP BY
columns are single-point, we can compute the number of GROUP BY groups.

Example: in the query:
SELECT ... FROM tbl
WHERE keypart1 IN (1,2,3) and keypart2 IN ('foo','bar')

Other things:
- Fixed cost calculation to more correctly count the number of blocks
  that may be read. The old code could use the total blocks in the file
  even if a range was available.

sql/opt_range.h

@@ -302,6 +302,7 @@ class SEL_ARG :public Sql_alloc
 {
   static int sel_cmp(Field *field, uchar *a, uchar *b, uint8 a_flag,
                      uint8 b_flag);
+  bool min_max_are_equal() const;
 public:
   uint8 min_flag,max_flag,maybe_flag;
   uint8 part;					// Which key part
@@ -401,6 +402,7 @@ public:
       return false;
     return true;
   }
+  int number_of_eq_groups(uint group_key_parts) const;
   inline void merge_flags(SEL_ARG *arg) { maybe_flag|=arg->maybe_flag; }
   inline void maybe_smaller() { maybe_flag=1; }
   /* Return true iff it's a single-point null interval */