feat(convert): publish table cell background colors

[willkg]

Fixes #7.

Authoring surface

A leading <!-- bg:COLOR --> HTML comment in a GFM table cell sets that cell's background in Confluence:

| Service | Status                     |
| ------- | -------------------------- |
| auth    | <!-- bg:light-green --> ok |
| billing | <!-- bg:light-red --> down |

An HTML comment won this over the alternatives (a delimiter-row extension, an in-cell {bg=…} token, a per-table directive) because it's invisible in a plain markdown preview and lives in the cell, so the color travels with the row when rows are moved, reordered, or copied — no coordinates to keep in sync.

COLOR is one of the 21 swatches from the Confluence editor's cell background picker, or a literal #rrggbb for anything off-palette. Atlassian doesn't document the palette, so the hexes were read off an editor-authored page; note grey is #b3bac5, and the picker's grey column runs white / light grey / grey rather than light/medium/bold. README has the full table.

Two cases are dropped with a warning rather than silently:

• an unknown color name;
• a marker that isn't the first thing in its cell. Confluence strips HTML comments on write, so a misplaced marker would otherwise vanish without a trace on the page.

Implementation

• tableCellBGTransformer (AST transformer, same shape as calloutTransformer) consumes the marker, trims the whitespace it leaves behind, resolves the color, and stashes the hex on the cell node.
• renderTableCell emits it as data-highlight-colour, otherwise reproducing goldmark's GFM cell output byte-for-byte — no existing golden changed. Cell alignment is untouched; that's still Table cell alignment is silently dropped when publishing #48.
• cellBGSwatches is the name→hex map, with the palette's provenance recorded in a comment so nobody re-derives it.

Testing

make test && make lint && make vet clean.

New regression case table-cell-colors/ covers swatch names, hex, header cells, a marker-only (empty colored) cell, case-insensitivity, the unknown-name warning, the misplaced-marker warning, and a non-bg comment passing through untouched.

Verified live against Confluence, since storage persisting an attribute doesn't prove the renderer honors it. Published markfluence's own output and read body-format=atlas_doc_format back: all 21 swatches reach ADF as a cell background with the exact hex (21 checked, 0 mismatched), plus an off-palette hex, a colored header cell, and a marker-only empty cell. Both warning cases published uncolored, and no alignment marks appeared anywhere — #48 is unchanged, not made worse.

Scope note

Per the discussion on #7: table width and column widths aren't needed (PR #50 made the default better), and column alignment is #48. Cell background color was the remaining piece, so this closes #7.