Add circleci compatible sharding

[raubitsj]

This seems like the easiest thing to do to get better performance:
https://circleci.com/docs/parallelism-faster-jobs

open question:
This can easily be done for a given environment (base circleci image), but does it make sense to use parallelism across different images. For example should we have a single functional test job for each machine type, then circleci parallelism that shards different python versions? different tensorflow versions. etc