feat: add composable low-level table components (TableRoot, TableHead, TableRow, TableHeaderCell, TableBodyCell) - #4977

Header rows no longer need a distinct row variant. A header row's visual identity (the shaded background) lives on TableHeaderCell, and its structural role is conveyed by placement inside TableHead, so TableRow variant narrows to default | selected | shaded.

This lets two more things go:
- The row-level header background only backed the grid underfill strip (header cells paint themselves); with it removed, an underfill header shows cells shaded to their tracks, consistent with body rows.
- grid-auto-rows only imposed a minimum row height on all-short rows; dropping it makes grid rows content-sized, matching auto mode and the existing Table, neither of which has a row-height floor.

…n tests

Replace the hand-rolled ResizeObserver in TableRoot with two useResizeObserver calls (the scroller and the table it wraps), matching how the rest of the codebase observes elements and getting a synchronous initial measure for free. The re-measure effect narrows to gridTemplateColumns — the one overflow transition an observer misses (grid tracks overflow without either observed box resizing); auto-layout content growth resizes the table box and is caught by the child observer.

The unit overflow-region test is rewritten to drive the transition through a column-template change (no ResizeObserver mock, which the shared hook's entry conversion would otherwise require). A new integ suite covers the two paths jsdom can't: keyboard-scrollability of the focusable region, a viewport resize starting/stopping the overflow (scroller observer), and auto-layout content growth starting it (table-box observer).

… data-attr

RowVariantContext's only runtime effect was the atomic body cell's shaded-background class. Selection already paints from the row's data-awsui-variant-selected attribute via a `> .cell` rule, and the shaded *border* already did too — only the shaded *background* went through the context. Move it to a matching `[data-awsui-variant-shaded] > .cell` rule and the context has no consumers, so delete it along with the per-row provider and both `value="default"` resets (TableRoot and the existing Table).

The existing Table keeps its TableContextProvider reset (defaultTableContext): its cells are the shared substrate and read the ambient column layout, so without it a Table nested in a grid-layout atomic cell would render grid roles/classes. That reset is load-bearing; the RowVariant one was not.

The variant tests asserted that atomic cells do NOT carry the existing Table's per-cell selection/shading classes — testing the absence of a class-based mechanism the atomic never used now that selection and shading both paint from the row's data-awsui-variant-* attribute. Drop those per-cell class-absence loops and the unused helper; each variant test now asserts the real contract (the correct data-awsui-variant-* hook on the <tr>, mutual exclusivity, and no aria-selected).

Also delete two tests that guarded removed/never-built behavior: the stray-cell test that existed only to check the deleted RowVariantContext default, and the grid roving-tabindex negative (a deferred feature the atomic never implemented).

…sed interface export

… props; selection wins over shading