An expression with an empty primary key, such as `dj.U() & T` or
`dj.U().aggr(T, n="count(*)")`, holds at most one row. Two operations on
such expressions failed:

- `len()` took the single-column branch and generated
  `count(DISTINCT )`, which is invalid SQL. It now returns 0 or 1 by
  checking whether the row exists.
- Restricting a `dj.U()` aggregation hit a bare assert in
  `Aggregation.make_sql`. Removing the assert alone would not have been
  enough: the MySQL branch emitted HAVING only together with GROUP BY,
  so the restriction would have been dropped without notice. MySQL now
  emits HAVING without GROUP BY, and PostgreSQL takes the subquery path
  whenever the aggregation is restricted, with or without grouping
  attributes.

Restriction conditions on an aggregation are now joined with str(), as
the base where_clause already does, so restricting any aggregation by
False no longer raises TypeError.

Closes #1581
Closes #1580