Required some changes to the inputters, to allow data to be passed in as
StringIO instead of a path to a file.

Re-enabling this test would require the following:
- An up-to-date config for a lightweight smoketest model
- An actual model trained from said config
- Updating the test to specify config and task_id

hydra-L-train-config.yml is a modified version of the training config,
with sentencepiece transform added.

server.json is required by the backend.

expected_files lists the expected directory structure, with model and
vocab files that are too large to include in this repo.