The second assert trips with EBU Tech 3280-style sample layout, where
the digital active area runs right up to the end of the line. It's not
actually a problem to run off the end of the line, because we'll just be
using the blanking data from the start of the next line (and there's
always another line after the last active line).

This gives better vertical resolution, at the cost of not dealing with
chroma phase errors. Thanks to Richard Russell for providing the specs
for the filter from the BBC's original Transform PAL implementation.

This only works with transform2d/3d at the moment, but in the future
it'd be possible to add a luma notch filter, giving a pal1d mode.

The core of this is effectively the same, once the setup is done; it
just needs to keep track of where 0H is for each line.

The count of lines in field 1 was wrong here, which meant that the Vsw
state didn't flip between the last line of field 1 and the first of
field 2.

Add a reference to Clarke for the intended characteristic.