Viewer: render figure exhibits as inline images

[EYH0602]

Background

Gap analysis vs the deployed ara-hub viewer (https://www.agenticresearch.sh/ara/AmberLJC/ara-paperbench/artifacts/speedrun/nanogpt-speedrun). Related: #46 (exhibit markdown styling).

Problem

figure > img rendering is an inert CSS stub (crates/ara-viewer/src/detail.rs:602-614, styles pre-skinned at crates/ara-viewer/public/styles.css:1126-1238). Figure-kind exhibits can only ship their content as a markdown body, so figures degrade to data tables — the nanogpt-speedrun artifact literally carries the note "Rendered as its data table -- this ARA carries no raster image for this figure."

The hub base64-inlines figure PNGs and renders them as bordered, padded <figure> blocks with mono captions.

Proposal

• Extend the exhibit payload so figure exhibits can carry an image reference (relative path resolved by ara serve, or inlined data URI for static export).
• Render ExhibitKind::Figure exhibits with an image as <figure><img><figcaption> (figcaption = description, see Viewer: render exhibit captions (description) in the result block #55).
• Fall back to the current markdown-body rendering when no image is present.

Acceptance

• Figure exhibits with an image render as inline images with captions
• Figure exhibits without an image fall back to current behaviour
• Images served correctly by ara serve and resolvable in static/embedded mode
• Tests for both branches