gstlal_inspiral_marginalize_likelihoods_online: change sleep to 10 minutes to allow for jobs to start

if the call-back to handle a new buffer raises an exception, the lock can
get released in the finally: clause allowing other threads waiting to go to
begin running before the finally: clause re-raises the exception.  if the
thread that starts running hits the assert related to un-processed buffers
then you get that exception instead of the one that really triggered the
failure, making debugging difficult.  this patch moves the call-back to the
very end of the try: clause so that there's no book-keeping left to take
care of in the event that it fails, which should ensure we get the correct
traceback displayed.

works around a bug in numpy < 1.7.  Thanks to Chris P. for tracking this
down and coming up with the fix.

patch moves receiver call into separate thread, uses a timed wait to pull
buffers from the thread.

We no longer rely on the chi column of sngl_inspiral to be
populated. NB: "chi" is different for different waveforms. Implemented
here is the effective spin usually associated with IMRPhenomB and
company. For the reduced spin chi (e.g. as in TaylorF2RedSpin), the
effective spin is not strictly correct but is probably close enough
for the purpose of splitting the bank and anyway is certainly better
than relying on a defunct sngl_inspiral column.

FIXME: should pass this information on from pbh_pipe

apparently some Macs don't have strndup().

This reverts commit a7529748.

Contrary to what I understood from the documentation, some of these flags are
not defined on other platforms on which this code is in common use.

an inspiral-specific sub-class of CoincParamsDistributions

patch OK'ed by Kipp

gstlal_inspiral_followups_from_gracedb: a snr / time ; snr / chisq plotting script to run on lvalert