perf(web): reduce first-load jank with chunked rendering and infinite scroll

[scoful]

Problem

The models table currently renders a very large DOM on first load.
On lower-spec devices and some server environments, this causes visible UI stalls, high CPU usage, and poor first-interaction responsiveness.

In addition, table column logic was duplicated across multiple places (header rendering, sorting, search text building), which made changes
error-prone and harder to maintain.

What this PR changes

• Switches the table to a lightweight shell + client-side data rendering.
• Loads model data from /api.json with /_api.json fallback.
• Introduces chunked rendering (CHUNK_SIZE = 100) to avoid blocking the main thread.
• Adds infinite scroll using IntersectionObserver.
• Adds a fallback path for environments without IntersectionObserver so all rows still render (no partial-data regression).
• Unifies table column definitions into a single source of truth (columns.ts) and uses it for:
  • header rendering
  • cell render order
  • sorting accessors
  • URL sort mapping
  • dynamic colSpan
• Replaces large index-based switch/case sort logic with typed accessor maps.
• Adds row hover/focus highlight for better readability while scanning large tables.

Why this should be merged

• Directly addresses the main UX pain point: first-load lag/freezing on large datasets.
• Preserves existing user-facing functionality: search, sort, URL state sync, copy model ID, keyboard shortcut.
• Reduces maintenance complexity and regression risk by removing duplicated column logic.

Validation

• cd packages/web && bun run build passes.
• Manual verification completed for:
  • progressive loading behavior
  • sort/search + URL updates
  • copy model ID interaction
  • non-IntersectionObserver fallback behavior

Scope

• No provider/model schema changes.
• No backend/function API behavior changes.