Previously we tried to do the whole template parse in one pass. This was
adequate for dealing with literals and interpolations because they
create a flat structure, but to parse the template control sequences we
need to be able to deal with nested template sequences.
As a building block towards this, we first do a pass of extracting the
template-level "tokens": literals, interpolations, control sequences.
We then pass over that sequence of tokens and parse it, which is then
simplified because the larger template atoms have already been produced.
