An assertion in readline() would fail as data was already in the
internal buffer even though the socket was in unbuffered read mode.
That case is now handled.  More importantly, read() has been fixed to
not over-recv() and leave newly recv()d data in the _fileobject buffer.

The max() vs min() issue in read() is now gone.  Neither was correct.
On bounded reads, always ask recv() for the exact amount of data we
still need.

Candidate for backporting to release25-maint along with r62627.