Add a way to test dvc repro on sample / test dataset

[nsorros]

At the moment dvc repro is not tested not guaranteed to run with changes. This is mitigated to a certain extent by tests but not fully. It would be ideal to have a way to run dvc repro on a test dataset so validate that it works before kicking a full run. This is intended to be used from the person that develops a PR as a sanity check similar to tests. Later on it could be added as Github check although a minor problem there is that dvc pull needs to run which fetches >100GB.

[nsorros]

One we were discussing with @pdan93 is to add an environment variable TEST=1 that dvc reads and uses as a param or flag in preprocess to create a small dataset say 1K examples or less. This can be enabled / disabled locally and in Github and it will not be enabled when running the final run in the cloud (or whenever its run)

[nsorros]

One way that works well in neural nets or models that use sgd variants to train and you can stop train early is to use a --dry-run flag which stops train after the first batch. We have effectively used this in #160