Keep Tile memory bounds semantics explicit

[maleadt]

Document tile memory bounds handling being explicit:

• ct.load and ct.store remain checked by default. Partial loads use their
  requested padding and partial stores clip to the destination.
• check_bounds=false is the explicit full-tile promise. It drops load padding,
  emits Tile IR's inbounds encoding, and requires Tile IR v13.4+.
• Julia @inbounds and --check-bounds continue to apply to Julia
  @boundscheck blocks, but do not change Tile memory padding, clipping, masks,
  or view attributes.

This applies consistently to direct TileArray accesses, eachtile windows
(including strided views), gather/scatter views, and scalar indexing wrappers.

Also included:

• Fix scalar not_int lowering for !::Bool and integer ~.
• Add codegen regressions for the explicit bounds boundary and GPU regressions
  for ragged tiles and unchecked padded loads.
• Document Tile memory bounds semantics, the Julia distinction, and the
  equivalent cuTile Python behavior.

[maleadt]

After looking into this more, I don't think we should couple Tile IR's check_bounds to Julia's @inbounds.

Julia's bounds checks are protective: eliding one should not change the result of a valid access. Tile IR's checked accesses instead define padding and clipping for partial edge tiles. A valid TiledView index can select such a partial tile, so making @inbounds tiles[i] discard its padding would change the value of a valid indexing operation.

The current pre-v13.4 fallback also loses padding_mode: the language layer drops it before the emitter discovers that it must fall back to a checked encoding.

Let's keep the boundary explicit:

• PR Resolve bounds-check markers and fold constant structured branches #279 handles actual Julia @boundscheck blocks, including @inbounds and --check-bounds.
• Tile loads and stores remain checked by default.
• check_bounds=false remains the explicit full-tile promise and requires Tile IR v13.4+.