These were not idle questions.  I wrote the patch, and I had to know what behavior is correct.

Here's the patch.  It fixes potential memory bug (unconsumed_tail sets to NULL in case of out of memory), resets the unconsumed_tail to b'' after EOF, updates unconsumed_tail and unused_data in flush().

I a little changed test for the previous case (here was O(N^2) for large data).  I checked it on non-fixed Python, bug was catched.