The test changes make them order-sensitive, which they should be.
It's important that our loader methods return date-sorted entries
just like Beancount itself would.
Building a string and loading it means Beancount can never cache any
load. It only caches top-level file loads because options in the
top-level file can change the semantics of included entries.
Instead use load_file as much as possible, and filter entries as
needed.
The old loading strategy didn't load options, which yielded some
spurious errors. It also created awkward duplication of plugin
information in the code as well as the books.
Implement a new loading strategy that works by reading one of the
"main files" under the books/ subdirectory and includes entries
for additional FYs beyond that.
This is still not ideal in a lot of ways. In particular, Beancount can't
cache any results, causing any load to be slower than it theoretically could
be. I expect more commits to follow. But some of them might require
restructuring the books, and that should happen separately.
