Try to optimize delay in LVP

A normal for loop like

    for (int i=5; i; --i) {
      delay_s();
    }

is unrolled until i=9. This is bad because a loop makes sense starting
at i=4. If the loop version (i>8) requires x byte, the unrolled version
(i=8) requires x+12 bytes. Due to compiler optimization this assembler
implementation requires x+4 bytes, although it just kind of forces
no-unroll of the loop.

So until we can tell the compiler to improve the jumps this does not
help to reduce the binary size.

rampinglight.c

@@ -509,9 +509,15 @@ int main(void) {
           // that the flashlight is still turned on but the battery is dying.
           // TODO If free space in flash, disable as many components as possible
           blink(3, FLASH_TIME/2);
-          delay_s();
-          delay_s();
-          delay_s();
+          uint8_t delay_seconds;
+          asm volatile (
+            "ldi %[i], 5"    "\n\t"
+            "0:"             "\n\t"
+            "rcall  delay_s" "\n\t"
+            "dec    %[i]"    "\n\t"
+            "brne   0b"      "\n\t"
+              : [i] "=&r" (delay_seconds)
+          );
         }
       } else if (voltage <= BAT_LOW) {
         blink(16, FLICKER_TIME);