Arithmetic operators don't represent SQL overloading

In SQL (Postgres, at least), arithmetic operators such `+`, `-`, `*`, and `/` are overloaded for date/time types. for example, the expression `timestamp '2001-09-29 03:00' - timestamp '2001-09-27 12:00` results in `interval '1 day 15:00:00'` (see https://www.postgresql.org/docs/9.2/interactive/functions-datetime.html).

However, Esqueleto defines `(-.) :: PersistField a => SqlExpr (Value a) -> SqlExpr (Value a) -> SqlExpr (Value a)`, which will generate the SQL needed to subtract timestamps, but gives the result the wrong type. Letting these operators be applied to just any type is incorrect anyway (e.g. you can't subtract strings in SQL), and forcing the argument and result types to all be the same is both unsound _and_ prevents valid queries from being written.

I'm not sure how this mess would best be represented in Esqueleto, but perhaps a set of additional operators, including one for `(TimeLike a, TimeLike b, IntervalLike c) => SqlExpr (Value a) -> SqlExpr (Value b) -> SqlExpr (Value c)` amongst several others. I'm not sure how many types would inhabit TimeLike, IntervalLike, etc. This isn't sounding super promising.

Also open to gross hacks that let me do the unsafe cast or something.

The payload names no files or tests; start by locating Esqueleto's definitions of (-.) and the related arithmetic operators, then compare their generated SQL and Haskell result types with PostgreSQL's date/time operator behavior. Done means valid date/time arithmetic can be expressed without an incorrect result type, while unsupported arbitrary operand combinations are not accepted.

Written by the indexing model from the issue text.