Suggest moving to csv

[akoumjian]

Moving to CSV as our base serialization format will have several benefits over the current h5.

Pros

• Much less disk usage (for the size of our files, we are not gaining anything from h5 compression)
• Much simpler file append mechanics
• File generation is deterministic, which means we can use hashing functions to quickly check for changes. This will let us efficiently do caching comparisons between all the different steps. Weirdly, h5 files are not bit deterministic on disk
• Easy visual inspection of serialized values on disk, useful for checking float precision and other errors

Cons

• We will lose the ability to use store.select for on the fly SQL - like analysis. This really only affects individual files larger than our machine RAM limits, otherwise we just read it into a dataframe in memory. We don't have too many of those since we break up the files by year-month (largest NSC is 2.2GB).
• we have some existing .h5 files that we like to reference / use. I've preserved ingesting by h5 in this PR, so we don't really lose that.

[akoumjian]

@moeyensj I've rebased off your latest schema changes. It's failing on the test, I believe due to reasons you already solved. Am I missing something in the build (reference to version of pyoorb?) that would fix the automated tests?