Crap, I'm working on something exactly like this :D Was about to upload it to Github just today, but still wanted to fix a few things.

Yours looks quite more finished, one thing that would be very useful is handling Python to Postgres type conversion, like PL/Python does. After that, the next step is caching I/O functions for the duration of the scan, which PL/Python also does.

Nice to see that the idea made sense for more than one person, hope Multicorn will rock!