fix(editor): let image captions be selected as document text

[horacioh]

Summary

Fixes #925: caption text could not be range-selected in edit mode, so the fragment actions built on selections (Copy Link, Comment → #blk-image[start:end]) were unreachable for captions.

Two independent causes, both in the image render path:

1. BlockSelectionWrapper wrapped the whole image subtree in contentEditable={false}. A ProseMirror view manages a single contenteditable range, so the caption (which is real inline content of the image block) became its own editing host: view.hasFocus() stayed false and prosemirror-view discarded every caption selection.
2. MediaContainer's outer wrapper was draggable=true, so caption mouse drags became native drags instead of text selection.

-<BlockSelectionWrapper editor={editor} block={block}>
+<BlockSelectionWrapper editor={editor} block={block} keepEditable={mediaType === 'image'}>

 // media-container.tsx: draggable moves off the wrapper (the caption's ancestor)
-<div className="…flex-col…" draggable={canAuthor} onDragStart={…} onDragEnd={…}>
+<div className="…flex-col…">
   <div {...mediaProps} contentEditable={false}
+       draggable={canAuthor} onDragStart={…} onDragEnd={…}>

The media surface itself stays contentEditable={false} and draggable, so click-to-select, side-menu drag, and reorder are unchanged.

Third change, a consequence of the first: Enter in a caption was handled by a React onKeyDown on the caption element, and the Enter keymap explicitly no-op'd for image blocks to defer to it. Now that the caption is part of the outer editing host, keydown targets div.ProseMirror and never reaches that React handler — Enter silently did nothing. The behavior moves into the keymap branch that used to be the no-op:

// KeyboardShortcutsExtension, handleEnter
if (blockContentType === 'image' && !(selection instanceof NodeSelection)) {
  const nextTextPos = TextSelection.near(doc.resolve(block.afterPos), 1).from
  nextTextPos > block.afterPos
    ? tr.setSelection(TextSelection.create(tr.doc, nextTextPos))   // move to next block
    : tr.insert(block.afterPos, blockNode.create(null, paragraph.create()))  // image is last
}

Tests

• media-container.test.tsx: new assertions that no caption ancestor is contenteditable=false or draggable=true, and that the media surface keeps both. Both failed before the fix. The three unit tests covering the removed React Enter handler are deleted — that behavior now lives in the keymap and is covered end-to-end instead.
• image-caption.e2e.ts: keyboard selection (asserting a non-empty TextSelection inside the image block, view.hasFocus(), and the formatting toolbar appearing), mouse drag selection, bold on a caption range, and Enter with the image as the last block.
• Full editor E2E suite passes (103 passed / 12 skipped). The 9 pre-existing unit failures in embed-editor, readonly-viewer-gallery and BlockManipulationExtension reproduce unchanged on the base branch.

docs/projects/image-caption-selection.md holds the investigation and the remaining phased test plan (media regression sweep, clipboard/SSR, mobile), plus links to the follow-up issues #926–#931.

Link to Devin session: https://app.devin.ai/sessions/923dddc749c646c8965bc2110a7d9ab4
Requested by: @horacioh

---