Refine the doubling logic in sample set

[duyhuynhdev]

Currently the doubling logic currently doesn't care how many events in a sample, as long as the sample's index is even it will be dropped.

• For example:
  If we have sample list as below:
  0 - 10 events
  1 - 1000000 events
  2 - 10 events
  3 - 10 events
  If we want to half that right now the logic would end up with
  0 - 10 events
  2 - 10- events

The important sample and a significant of event has been removed and its could not show the bottleneck (see #204)

So that the logic for doubling should be changed. Instead of dropping every second row we can merge them instead by

1. Adding the event counts
   For example :
   0 - 10 events
   1 - 1000000 events
   Would become
   1 - 1000010 events

2. Picking the stack of the one which is bigger

[duyhuynhdev]

Another idea from @brendanheywood , see below:

0 - 10 events A
1 - 100 events B
2 - 100 events C
3 - 10 events D
4 - 10 events E
5 - 10 events F

if we halve from 6 to 3 then we break it into 3 groups of events. The total event count is 240 so we want 3 groups of 80

0 - 10 events A
1 - 70 events B

1 - 30 events B
2 - 50 events C

2 - 50 events C
3 - 10 events D
4 - 10 events E
5 - 10 events F

then for each group we combine and pick the large event
0 - 80 events B

1 - 80 events C

2 - 80 events C
there will be rounding errors no matter what we do