Support New Segmenter

[bitsofbits]

General Support

• Pass last course and speed and first course, speed, lat, lon through pipeline
  so that they are available on segments imported from previous day.
• Collect identity stats for each segment (shipnames, callsigns, etc).
• Generate two sets of segment output -- one that is compatible with previous
  pipeline and one that is for the new stitcher and dump to different output files.
• Keep segments alive, even with no messages, until closed
• Factor out core of segmenter implements so that it's easier to call externally for testing.

Software updates.

• Use pipe-tools 3.0: this fixed a bunch of versioning problems and
  needed to be done eventually.
• Make tests run under Python 3 and Python 2. Doesn't necessarily mean whole
  pipeline will run under Python 3 yet.

Reduce Memory Usage

See Issue #100

Improves memory usage / performance in three ways:

• Filters out noise segments before grouping. The new segmenter generates more
  noise segments so this results in a large improvement in memory footprint.'
• Cogroup segments and messages rather than passing segments in as a side argument.
  This means each worker machine only gets the segments related to the messages it is
  processing further reducing memory footprint and slightly speeding operations.
• Use shards when writing segments as well as when writing messages. Improves the write time
  segments, and may reduce per worker memory usage.

Net effect is that I can run over 1 month with 8192 MiB per worker, and perhaps less (that's the minimum I've tried), when previously I was still getting crashes with 64,000 MiB per worker!

[enriquetuya]

@bitsofbits check my questions and comments below.
Also, we need to add an entry to the CHANGES.log stating the changes.