Make model generation more (maybe fully?) deterministic (#871)

The generation was not very deterministic, and the reproducibility test
wasn't catching it.  Dropping and recreating the database between cycles
exposed a lot of issues (including still embedding ids).

 * Don't embed id in schema object values. Put a dummy in instead.
  (Will this be a problem?)
* Sort groups of potentially overlapping functions by their
  CallableSignatures
* Inside overlapping groups, sort by edgeql_signature in addition
  to their "generality_key"
* Don't use base64 of ids to generate Tuple names. Instead use
  a base64 of name. This produces kind of horrific names; at some
  point we should try to safely produce something based on the
  actual name.