feat(render): rewrite VL→matplotlib as a staged tree-walking interpreter

[Roy-Kid]

What

Rewrites the Vega-Lite → matplotlib renderer from an ad-hoc, specs.py-coupled
prototype into a small tree-walking interpreter (molplot.vlmpl) whose
source language is Vega-Lite and target machine is matplotlib.

Four passes, each a classic compiler stage:

1. model.py — normalize the raw spec into an immutable, field-name-agnostic
   AST (Unit/Channel/Scale); resolves layer inheritance and mark shape.
2. scales.py — bind scales: colour maps, per-series numeric maps, and
   positional axis config.
3. marks.py — dispatch each mark to an encoder (MARK_ENCODERS registry);
   encoders read typed channels (enc["x"].field), never hardcoded field names.
4. axes.py — finalize titles, positional scales, and legend.

render.py is now a thin shim re-exporting molplot.vlmpl.render; the public
molplot.render(spec, *, preset, mode, ax, apply_style) API is unchanged.

Why — two parity gaps the prototype had

• Positional scales were dropped. specs.py emits scale.type: "log" and
  scale.domain, but the old renderer never applied them — a web log axis
  rendered linear in matplotlib, and x_domain/y_domain were ignored. The
  axes pass now applies both.
• Per-layer transform filters were ignored. The line spec's marker layer
  is gated by transform: [{filter: {oneOf: markerKeys}}]; the old renderer
  drew marker points on every line chart. Filters are now honoured, and the
  detail channel splits series that share a legend colour.

Tests

test_vlmpl.py adds coverage for the fixed behaviour (log axis, domain limits,
no stray markers, detail splitting). Full suite: 33 passed (27 existing + 6
new).

Note for reviewers: the local pre-push hook runs cd python && pytest
against the editable-installed molplot, which points at your main checkout —
so it can show the new tests failing against old code if you have a stale
install. pip install -e ./python on this branch (as CI does) → all green.