TODO

- Some nuisance issues related to thread termination:
   - if a C++ exception is thrown, it takes a long time for the core dump to get written
     (I suspect this would be fixed by properly joining threads after they exit)
   - if a C++ exception is thrown it crashes the python interpreter, since cython is
     not trapping the exception and converting it to a python exception (not sure why,
     this works in toy examples)
   - if C++ processors are being run through the python interface, then control-C
     doesn't abort the processing, since it's being "swallowed" by the python interpreter
     (an easy fix would be to add a 1-sec timeout to assembler.wait_for_end() and
     loop it)
   - it appears that the vdif_assembler destructor doesn't always get called at the end
     of ./run-vdif-assembler (?)

- Reinstate a few loose ends from v1 (most importantly the RFI mask)

- Performance testing on moose in a "fully loaded" case with network capture and writing
  to disk, at the same time as a nontrivial C++ processor such as an FRB search.

- Quantify packet loss.  Is there any benefit to using high-tech libraries like pfring or dkpk?
  Related: pin threads to cores?  Use hugepages?

- Why is the assembler so much slower on my desktop than on moose?  Is the L3 cache
  becoming a bottleneck?

- assembler_nerve_center::cond_rbuf_* should be arrays of length rbuf_size.
  (I think it's OK to keep cond_abuf_* as a scalar, right?)

- Test concurrency corner cases

- Can we define a single templated structure which combines the various thread-safe
  ring buffers defined in the code?  (rbuf, abuf, unit_testing_buffer)

- We may eventually want a little more asynchronous functionality, so that the
  assembler can run as a "remote-controlled" daemon

- 256-bit kernels

- To avoid the page faulting cost of glibc malloc/free, I'm currently using some
  homegrown memory management which never returns buffers to the operating system
  until the managing thread exits.  Is there a better solution?

- Cleanup:
    - Properly destroy mutexes/conditionals, and joining threads
    - Use a public/private split in classes (currently I'm not bothering in most cases)
    - Makefile should warn if . and $(LIBDIR) aren't in LD_LIBRARY_PATH (in this order)
    - likewise if $(PYDIR) isn't in PYTHONPATH
    - Numerous small FIXME's throughout the code