Set metadata schemas

[hyanwong]

Defaults to a "struct" type unless a schema already exists (or if there is a null schema that can be interpreted as JSON). If so, the schema is modified to allow the mean time and variance in time (mn & vr) to be saved as extra values.

To maintain backwards-compatibility, I have used the top-level keys "mn" and "vr" for the mean and variance in times (the description field in the schema should make it clear what these are).

Fixes #302

[hyanwong]

Also fixes #203

[benjeffery]

Looks good @hyanwong, we really should get the metadata updating methods into tskit proper, but I think they'll have a similar behaviour to what you've done here.
How much do we care about backwards compatibility here? The two character names suck, and with struct there is no cost to longer, more descriptive names.

[hyanwong]

Thanks @benjeffery. I agree that the 2 character names are a bit crap. I'm not that concerned about backwards compatibility, TBH, but at the moment, tsinfer sets JSON metadata on nodes, and with this PR to tsdate, we don't change the codec when adding the Tsdate-specific metadata. So long names may kill us, storage-wise, when running tsdate on the current tsinferred trees.

If we address tskit-dev/tsinfer#416 then this ceases to be an issue, as tsinfer would output struct metadata for nodes too. I'm inclined to merge this PR (which doesn't really change behaviour much), but open another to say that we should switch to a more legible naming scheme when tsinfer switches to struct?

I assume that you think struct on node metadata in tsinfer is a good call?

[hyanwong]

Obsoleted by #383 . We can revisit using struct metadata later.