- recognizing that nothing uses the trigger class defined in
  snglinspiraltable except the trigger generator and its coincidence
  machinery, flatten the inheritence tree by one to improve attribute
  look-up performance

- move performance-critical methods to the C layer

- forgot a line from e3b38b0d

- use .template_id to check for template equality

- the coincidence code will now rely on the .template_id attribute to
  determine which template each trigger has come from.  this patch adds an
  assert to confirm that that is safe to do.

- a thinko was causing it to include one too many segments below the bound
  than required to satisfy the criterion.  it wouldn't lead to incorrect
  results, just a small slow-down.  ligo.segments now provides an
  implementation of this function, but that version of ligo.segments is not
  yet in production so we need to retain this implementation here for the
  time being.

- avoid an attribute look-up inside lower_bound_in_seglist()

This reverts commit 62952454.

ll_simplify_and_cluster.sql: delete high far events (above 1e-4) from the databases to keep the size manageable.

inspiral.py: choose only the best coinc for gracedb uploads, and fix performance issue of generating an xml doc when there are no uploads to do

gstlal_ll_inspiral_daily_page: fix some comments and comment out plot background code, which isn't working

- add snr_min cut to remove sub-threshold triggers in case itac starts
  producing them

- use new "event_collector" feature of snglcoinc to identify singles
  suitable for use in background models.  this should fix the problem of
  the sum-of-SNR^2 cut introducing a bias into the noise model PDFs.  in
  any case it decouples background collection from other aspects of
  candidate construction, allowing for the selection criterion to be
  adjusted more easily, with less thought required.