better errors in the expression package

[mccanne]

We should revamp and document all errors that come from the expr package.

In particular, error should be structured, detailed, and stacked so that debugging
is easier. (Related issue: #4019)

While we're doing this, we should change operators where NULL doesn't make sense
to create an error value rather than a zero value. Currently, NULL+1 is 1. In SQL,
NULL+1 is NULL but this creates headaches because it's vague SQL doesn't have first-class errors. (Related issue #4018)