Add MLPerf logging

[hanlint]

This PR contributes a callback MLPerfCallback, which will create a submission directory and results file that is compliant with the MLPerf rules for Training v1.1 (e.g. it passes the mlperf logging's package checker).

Upon usage, a submission folder structure will be created with the root_folder as the base and the following directories:

    root_folder/
        results/
            [system_name]/
                [benchmark]/
                    results_0.txt
                    results_1.txt
                    ...
        systems/
            [system_name].json

For each training run, a results file will be created, e.g.

:::MLLOG {"namespace": "", "time_ms": 1648183840286, "event_type": "INTERVAL_START", "key": "cache_clear", "value": null, "metadata": {"file": "/composer/composer/callbacks/mlperf.py", "lineno": 109}}
:::MLLOG {"namespace": "", "time_ms": 1648183840438, "event_type": "INTERVAL_START", "key": "init_start", "value": null, "metadata": {"file": "/composer/composer/callbacks/mlperf.py", "lineno": 110}}
...

The entire directory can then be submitted to mlperf, and checked with the package_checker in https://github.com/mlcommons/logging.

Currently this callback only supports the OPEN division benchmark.

This PR is gated by:

• PyPI package for mllogger mlcommons/logging#218
• Callbacks need access to computed metrics #830

TODO:

• generate and validate ResNet-50 run on 8x A100 GPUs.
• create Hparams object

[ravi-mosaicml]

Overall looks good; left some comments for now. Will re-review once the open TODOs are addressed.

[ravi-mosaicml]

Really close! Just a few nits, plus one comment re: binding the train dataloader and evaluator to the state after Event.INIT runs. Rational is that these parameters will not need be specified in Trainer.init(...) after #948 lands (and algorithms / callbacks are already OK with no dataloaders or evaluators during Event.INIT).

[ravi-mosaicml]

LGTM 💯! See comments