Use a run_task abstraction for eval-tests.py

[gregtatum]

Instead of testing at the script level, this tests at the run-task level, which I feel is the integration test level.

    run_task(
        task_name="evaluate-quantized-sacrebleu-wmt09-en-ru",
        script="pipeline/eval/eval-quantized.sh",
        work_dir=test_data_dir.path,
        fetches_dir=test_data_dir.path,
        env={
            # This is where the marian_decoder_args will get stored.
            "TEST_ARTIFACTS": test_data_dir.path,
            # Replace marian with the one in the fixtures path.
            "BMT_MARIAN": fixtures_path,
            # This is included via the poetry install
            "COMPRESSION_CMD": "zstd",
        },
    )

I'm keeping this as draft because it contains work from #364 that I would like to land first. I also wanted to get a CI run in. The most relevant commit here is ecd2793

Edit: This is ready for review.

[gregtatum]

So this test was really useful in #401 rewriting eval.sh to eval.py and simplifying the task graph, as I could re-run the test to make sure I wasn't breaking things.

[gregtatum]

Here is an example of an existing test using run_task that I'll put up for review later:

@pytest.mark.parametrize(
    "params",
    [
        ("mtdata", "Neulab-tedtalks_test-1-eng-rus"),
        ("opus", "ELRC-3075-wikipedia_health_v1"),
        ("flores", "dev"),
        ("sacrebleu", "wmt19"),
    ],
)
def test_basic_corpus_import(params, data_dir):
    importer, dataset = params

    data_dir.run_task(
        f"dataset-{importer}-{dataset}-en-ru",
        env={"COMPRESSION_CMD": COMPRESSION_CMD, "ARTIFACT_EXT": ARTIFACT_EXT},
    )

    prefix = data_dir.join(f"artifacts/{dataset}")
    output_src = f"{prefix}.ru.gz"
    output_trg = f"{prefix}.en.gz"

    assert os.path.exists(output_src)
    assert os.path.exists(output_trg)
    assert len(read_lines(output_src)) > 0
    assert len(read_lines(output_trg)) > 0

[eu9ene]

Looks good overall! I have a minor remark about asserts. Running tasks is a great way to integrate a bit deeper and improve coverage.