Increase benchmark accuracy by using a result based on the last meassured speed after benchmark values changed by less than 0.1% after kernel repeats

Goal is a "what you see is what you get" value compared to: "singlehash -a 3 ?b?b?b?b?b?b?b" -- both with the same fixed -u and -n values
As a positive side-effect, this decreases total benchmark runtime
Add speed_cnt_total and speed_ms_total as a preparation to get rid of SPEED_MAXAGE which produces 0H/s display on very slow-hash types
Replace some floats with double which can (theoretically) become really big

include/types.h

@@ -955,9 +955,18 @@ struct __hc_device_param
   uint    exec_pos;
   double  exec_ms[EXEC_CACHE];
 
+  // this is "average" speed, we'll use this for benchmark and final status screen
+
+  u64     speed_cnt_total;
+  double  speed_ms_total;
+
+  // this is "current" speed
+
   uint    speed_pos;
   u64     speed_cnt[SPEED_CACHE];
-  float   speed_ms[SPEED_CACHE];
+  double  speed_ms[SPEED_CACHE];
+
+  // speed_rec is what additionally limits the "current" speed in time, not array elements
 
   hc_timer_t speed_rec[SPEED_CACHE];
 
@@ -1291,7 +1300,7 @@ typedef struct
   hc_timer_t timer_running;         // timer on current dict
   hc_timer_t timer_paused;          // timer on current dict
 
-  float   ms_paused;                // timer on current dict
+  double  ms_paused;                // timer on current dict
 
   /**
     * hash_info and username