In particular, lower all control-flow-involving expressions (such as boolops,
list comprehensions, if expressions) into actual control flow.  This requires
unfolding the AST into a flat structure, since the control-flow-expression can
be nested inside other expressions, including other control-flow-expressions.

The IRGenerator was doing this before, but was leading to duplication because
it's very similar to what the CFG generator already has to do.

Doing this earlier in the pipeline means that more analysis passes can work on
the lowered expressions, rather than having to be taught how they work.  In
particular, the name analysis does not need to have to special case the fact
that list comprehensions may set names -- but aren't guaranteed to if they don't
get executed or the iterator is empty.  I tried implementing list comprehensions
without doing this, and a bunch of the work was in reimplementing the logic that
the analyzers already implement.

As a side benefit, deopts become much easier since the AST is already unfolded;
this change gets rid of the messy fake-variable-setting way that the IRGenerator
was unfolding the AST to get this same effect.