add smoketest for basic training runs

[JamesKunstle]

This PR adds a singular smoketest that will run "training" on the sample data that we have in this repo. This iteration of the training loop has minimal features- no LoRA, no CPU offloading, no flash-attention. (i.e. minimal external dependencies).

It does require four GPUs, however, as a default. Currently it takes about six minutes to finish on a 4xL40s machine because the sample dataset has been truncated by only accepting samples that have a sequence length of <160 tok.

It executes:

1. Preprocess sample data, store in cache.
2. Load model, store in cache.
3. Start 4x accelerator training, set up four cards.
4. run training on 1 epoch (roughly 5 minutes to finish on tested hardware)
5. save full state checkpoint to disk.
6. exit loop, clean up cache and checkpoint.

It includes boilerplate for future feature-coverage tests that include stuff like LoRA or CPU offloading.

The intention is that this test (and following tests like it) should be run less frequently than traditional unit tests, linting, and static code analysis (hence the @pytest.mark.slow decorator) but should still give a clear indication, later in the PR review process, that our features will run to completion without obvious problems.