v0.3.0 — loading and error states for every artifact, plus image and video

Every artifact now has three presentations instead of one, and media joins the kinds. The additions
are backward compatible: existing kinds, segmenters, and the streaming engine are unchanged, and the
new loading state sits in front of behaviour that previously started at streaming.

• Artifacts have three presentations, not one. A new loading lifecycle state covers every artifact
  that exists on the canvas with nothing renderable yet, and every kind now compiles a loading and
  an error presentation alongside its content: skeleton bars for text-like kinds, a reserved box for
  media, and failures that show their message beside the typed code. Streamed artifacts stay in
  loading until their segmenter commits a first character, because a chunk can arrive that commits
  nothing.

• image and video artifact kinds. Media is atomic — no prefix of the bytes is a smaller picture
  — so MediaIngestion replaces segmentation with load, resolve, or fail, reporting determinate
  progress when a length is known and reserving the frame from the intrinsic aspect ratio so layout
  does not jump on decode. Decoding stays with the host; the engine owns the lifecycle, pin, and
  revision.

• Published to npm. The libraries are Apache-2.0 and live under @dopejs/canvas-*, built from the
  v0.2.0 tree with npm provenance attesting the GitHub Actions build. Installing is now ordinary:
  pnpm add @dopejs/canvas-core.