Overhaul Transform API

[b5]

After numerous design discussions, we've arrived at a new API for running transforms. This commit makes a number of breaking changes at once:

• download(ctx) and transform(ds,ctx) special functions are removed
• a new dataset module is loaded by default into all transform scripts, the module is scoped to the dataset the transform script is bound to
• a new module function dataset.latest() takes no arguments and returns the HEAD dataset version, returns an empty dataset if no history exists
• a new module function dataset.commit(ds) requires a dataset argument and must be called by the user exactly once in a transform script to update dataset. The dataset passed to commit is passed directly to save, without merging into the prior version
• calling dataset.commit halts the program
• calling dataset.commit zero times produces a warning, but is not an error (will error with no changes in practice)
• starlark context module is removed
• config is an immutable dict-like type replaces ctx.get_config(key) with config.get(key[, default]). config is loaded into the transform script by default. config may inherit more dictionary-type properties in the future.
• secrets is an immutable dict-like type, replaces ctx.get_sercret(key) with secrets.get(key[, default]). secrets is loaded into the transform script by default. secrets may inherit more dictionary-type properties in the future.
• dataset.TransformStep.Category is no longer used
• starlark steps are no longer frozen after execution, allowing symbol manipulation across steps

An example conversion from old to new:

OLD:

load("http.star", "http")
---
def download(ctx):
  csvDownloadUrl = ctx.get_config("url")
  return http.get(csvDownloadUrl).body()
---
def transform(ds, ctx):
  # ctx.download is whatever download() returned
  csv = ctx.download
  # set the dataset body
  ds.set_body(csv, parse_as='csv')

NEW:

load("http.star", "http")
ds = dataset.latest()
---
csvDownloadUrl = config.get("url")
res = http.get(csvDownloadUrl).body()
---
ds.body = res
dataset.commit(ds)

BREAKING CHANGES: We've overhauled the starlark API to use a cell-based approach. All transform scripts will require updating

[b5]

@dustmop, after a multi-hour rebase nightmare, I've got most of the changes we discussed re-worked atop the new dataframe api. tests in the cmd package aren't passing, but while I get those fixed up I'd love a review.