Developer guide says to use spans for consecutive years.

The code has been cleaned up, and tests added.

It's hard to explicitly provide correct keys without understanding how
the Butler dimensions system works in detail. Moving key constraints
to automated (if simple-minded) code greatly reduces the burden
on callers.

Each test should have its own collection for isolation, but creating a
completely new repository each time is impractical.

This change lets tests avoid using Numpy arrays, whose nonstandard
__eq__ behavior makes them poor test objects.

This approach is not object-oriented, but making the Datastore interface
support possibile child datastores is a nontrivial fix.