feat: Add "auto" unit to timeunit transform

[paulrosenzweig]

This pull request adds a new "auto" unit option to the TimeUnit transform. When set to "auto", the transform will automatically choose the appropriate time unit based on the input data.

I wanted to try and solve the problem discussed in #1310: plotting discretized timeseries data without knowing the grain.

Here's a demo Observable notebook. It patches the timeunit transform and uses DuckDB to produce differently grouped time data.

Problem/Solution

Currently, you can discretize data into months using this transform.

{"type": "timeunit", "field": "date", "units": ["year", "month"]}

But, if your data is already grouped by month, there's no way to infer that from the data.
This PR adds an "auto" unit option that would work with this example data.

[{"date": "2023-01-01"}, {"date": "2023-02-01"}, {"date": "2023-03-01"}]

{"type": "timeunit", "field": "date", "units": ["auto"]}

Alternate approaches

In an issue comment, @jheer suggested creating a new discretizing time scale that acts like a band scale. I believe that discussion happened before the TimeUnit transform was added. Enhancing TimeUnit seemed more natural fit than adding a scale, but I'd love to hear any other thoughts! I meant this PR mostly as a discussion starter with some running code.

Questions

• Is the API right?
  I'm uncertain about the "units": ["auto"] parameter. Would a separate param be better? e.g. "inferUnits": true or similar?
• Should I pull in the dependency?
  I created a small npm package that does the work to determine the likely unit/step. It felt like a separate concern that's not specific to vega, but it's very small. Should I pull that into vega-time as part of this PR?
  The main reason to leave it separate IMO is to add tests that would be annoying to have in Vega. I want to run the tests with Node booted into different timezones, but that's annoying requirement to put on a bigger project.