Raise error if data type domain differs between R and database

[krlmlr]

It's the only safe way. We might need to make this an option, though, to support legacy user code.

• Embedded NUL
• 2⁻³¹ vs. NA
• Number out of range

[krlmlr]

With ALTREP, we can achieve this:

• fetch the data in a format closer to the database library
• create R data structures on the fly, as necessary (e.g. postpone string allocation)
• offer different representations if a conversion into the default format is impossible

[krlmlr]

For now, let's convert types in SQL. Maybe in a future rewrite.