historical: Fix the rate ordering with two dates.

Darnit, I wrote the tests first, and I wrote them right, and then I
mixed up the ordering in the code, and somehow I convinced myself
the code was the right and the tests were wrong. But no, I had the
tests right, this is really what we want. This gets the output to
follow the examples from our bookkeeping documentation.

historical: Always format rates with the same precision.

When we format a rate as a price, we don't know how much precision
is "enough" to do the conversion, because we don't know what's
being converted to. As a result, we may (=will almost certainly)
end up formatting the rate with different precision on the cost
date vs. the price date, and that causes Beancount/Ledger to fail
to make the connection between them.

Using a constant of 6 is enough to make the current test for
"enough" precision pass, so just do that for now. This might need
further refinement in the future.

historical: Beancount can handle commas in amounts.

And having it looks nicer, is more consistent with our historical
books, is less code for me, and is no more trouble for the user.

tests: Historical tests use more flexible rate matching.

This lets the tests be more flexible about how much precision is used in
rates when appropriate, and makes them ready to parametrize for
Beancount.

historical: Introduce --output-format option.

Beancount output will be added here.

tests: Stop calling fixtures directly.

Avoid deprecation warnings from pytest.

historical: Add a setting to denominate Ledger conversions.

This makes conversion output easier to add to ledgers directly.