feat: carry variable bindings through the write path

[awdr74100]

Completes the half #164 deliberately left out. Reading a style, editing it and writing it back
still replaced the token reference with the frozen literal beside it, so the style silently
stopped tracking its variable — the exact failure the read fix made visible.

What now round-trips

set_fills · set_strokes · set_effects · set_layout_grids · set_text_range (per-run
fills) · create_paint_style / update_paint_style · create_effect_style /
update_effect_style · create_grid_style — each takes the same boundVariables
({ field: variableId }) the read side reports.

Two measured facts decided the shape

Probed against a live file rather than inferred:

what was tried	result
assigning a paint / effect / grid literal carrying boundVariables	accepted, and the value really is driven by the variable
the same for a gradient stop	accepted — and it is the only route, setBoundVariableForPaint takes a SolidPaint
official setter, wrong-typed variable	throws (can only bind color variables to color)
literal, wrong-typed variable	🔴 accepted silently, renders white
literal, id matching no variable	🔴 accepted silently, renders white
literal, non-bindable field name	Figma rejects it

So pass-through alone was not acceptable: Figma will quietly store a broken binding. Bindings go
through the official setBoundVariableFor* setters wherever one exists, which reject both bad
cases; the gradient stop, which has no setter, is type-checked here instead. Every id resolves
before anything is assigned, so a stale binding throws with the id named rather than
half-writing a node — or, for create_*, leaving a wrecked style behind in the design-system
panel to be published to the library from there.

Unbinding needs no separate path: Figma replaces the whole array, so a paint or effect written
back without its boundVariables clears what it was bound to. That is now stated in the tool
descriptions rather than left to be discovered.

Also fixed on the way

The gradient-stop object inside the paint schema was strict, so a bound stop was stripped
server-side before it ever reached the plugin. The layout grid was strict too, and carried inline
in two tools whose copies had already drifted apart — extracted to grid-schema.ts alongside the
paint and effect ones it should have matched from the start.

Verification

• 19 mutations of the new logic, each confirmed to fail at least one test. Three survived the
  first pass — no handler-level test asserted that set_fills, update_paint_style or
  update_effect_style actually applied bindings, and three others were caught only incidentally.
  Tests added; all 19 now caught.
• Live, end to end, on a real file (built plugin, real relay): a bound solid paint on a node
  and via create_paint_style; a shadow with color and radius bound via
  create_effect_style (read back as radius: 40, the variable's value, not the 4 sent); a
  gradient stop bound while its sibling stayed unbound; a layout grid's gutterSize;
  update_paint_style clearing the binding when omitted and restoring it when re-sent; and an
  unresolvable id failing with create_paint_style: variable VariableID:9999:9999 not found while
  creating nothing.
• The recorded plugin contract moves by 15 arguments, all of the silent class.
  MIN_PLUGIN_VERSION already stands at 0.5.0, the release these ship in.
• Every commit builds and passes the full suite on its own.