Fix memory spike when compacting overwritten points

[jwilder]

• Rebased/mergable
• Tests pass
• CHANGELOG.md updated

If a large series contains a point that is overwritten, the compactor
would load the whole series into RAM during a full compaction and deduplicate
the points. If the series was large, it could cause very large RAM spikes and OOMs.

This change reworks the compactor to merge blocks more incrementally
similar to the fix done in #6556. It also replaces the deduplication method with
a sorted merge approache used in #6556.

To reproduce the issue, I wrote 25M points to a single series and then overwrote the first point. When a full compaction runs, RSS spikes up quite drastically.

There was also a bug discovered in the newly introduced *Values.Merge funcs where the values being merged could end up unsorted. This is now fixed.

Fixes #6557

Some before and after stats:

Before

$ ps -o rss,vsz $(pgrep influxd)
   RSS      VSZ
3148776 145316856

[tsm1] 2016/05/05 21:53:49 compacted full 3 files into 1 files in 48.80575863s

After

$ ps -o rss,vsz $(pgrep influxd)
   RSS      VSZ
 32584 145292860

[tsm1] 2016/05/05 21:48:00 compacted full 4 files into 1 files in 8.040986536s

After compaction completed, RSS decreased from 3.1GB to 32MB and compaction time decreased from 49s to 8s.

[jwilder]

@benbjohnson @e-dard

[benbjohnson]

👍