taskcluster tests run with a different version of taskgraph than decision tasks/training runs

[bhearsum]

I noticed this in https://github.com/mozilla/firefox-translations-training/pull/377 today, where I've pulled in a new version of taskgraph to pick up a helper function, but had the test task fail with an error that indicated an older version was being used.

Whenever taskgraph runs in CI, we use the version from https://github.com/mozilla/firefox-translations-training/blob/main/taskcluster/requirements.txt. When we run unit tests, we use whatever is in https://github.com/mozilla/firefox-translations-training/blob/main/poetry.lock.

In order to ensure that we always test with what we run on our infrastructure, we may want to consider moving Taskcluster unit testing to a separate target/task, and using the requirements from requirements.txt instead of poetry.lock.