…2) get all ancestors of a set of tiles

…default way to tile images with a large FoV

…ges with large angular size

…um things like TilingMethod.TOAST actually equals TilingMethod.TOAST. Also added assertion for the resulting imageset's projection type after creation by tests.

…rly cases

…180 degrees of longitude

In one test image I have, hipsgen produces a "BAD PIXEL CUT" warning and
doesn't emit this information. (I think it's because the file has an
integer data type.) Avoid the crash if this is the case.

…de the other if it exists

As noted in the comments, my DASCH sample file that crosses the pole has
WCS headers that seem to have another problem beyond their specification
of the TPV polynomial distortions. I need to set LONPOLE explicitly to
get the correct orientation on the sky.

The "initiate readiness state" algorithm repeatedly tests whether
positions are in the tiles_to_process set. If large numbers of tiles
(tens of millions) are being processed, the algorithm basically fails
because such tests have O(n) scaling with lists. With a set, things
become tractable.

There may well be further room for improvement here, but this one change
makes an enormous difference for big problems.

In processing a FITS-to-TOAST dataset going down to level 14, with about
24 million tiles to process, I discovered that the tile filter function
was a non-trivial bottleneck -- it took about 10 minutes just to count
the tiles that needed to be processed! Some profiling showed that most
of the time was actually in the Numpy min/max functions. Here we
Cythonize the function and avoid the Numpy functions. My measurements
indicate a speedup of a factor of around 20.