Make the SPM country tests unit tests, with two integration tests against the installed US model #4135

Problem

tests/unit/test_country_spm.py took about 14 of the 41 minutes of the pull request workflow's Test job on #4039. make test runs tests/unit in one process under coverage run --branch.

37 of the file's 63 tests ran the installed policyengine-us model through two function-scoped fixtures:

• real_canonical_country reloaded and revalidated the 26.6 MB SPM forecast for every test (about 0.65 s each), although policyengine-us already caches it.
• real_http_client built a new Flask app, calculation service, response cache and database row for every test.

Most of those cases repeated one error-to-response mapping for each combination of settings, which stand-in tests can check in milliseconds. One case ran /us/calculate-full, which calculates every variable in the model; it took 48.6 s locally, and 95.0 s when repeated in the same process.

Goal

• The unit suite holds only fast unit tests, with stand-in simulations and no real model or forecast.
• The checks that need the installed model are two integration tests, sharing one app and service for the session.
• /us/calculate-full is checked with a stand-in of the route's behaviour.