Other parts of the code (such as the StreamChangeCache) assume that there will
not be multiple changes with the same stream id.

turns out that as of py37/PEP424, if you let a StopIteration bubble out of a
generator (as opposed to implicitly raising one by returning from said
generator), it gets turned into a RuntimeError.

Fix it by having `next` return None rather than raising StopIteration, and
handling that.

It's pointless but valid to invalidate the same entity twice at the same stream
id.

Argh. It turns out that if you introduce sanity-checks, you discover the places
in which we are not adhering to them.