Implement / revive convergence test

[justheuristic]

It would be very useful to automatically test that some new feature haven't broken model training.

Please implement a test that trains a simple model on synthetic dataset or MNIST (preferred) and test that it converges after a certain number of steps (i.e. loss is below some threshold).

Back in the day, Maxim implemented a basic MNIST training example that could come in handy.
You can find it here: https://gist.github.com/justheuristic/01d5ffe9c534d90e40badff35653ba7d

Recommendations:

• Test should be easy to run, preferably a single command like this test;
• Make sure tests can be run without GPU so that we can run them automatically for pull-requests;
• When applicable, use test_utils module;

[justheuristic]

Merged into #2