[GenericOp] Properly handle slice layouts in MakeRange fusion and unrolled LLVM lowering

[alexbaden]

Fixes two bugs which allow all previously failing reduction tests to pass under TileAndFuse.

• MakeRangeOp fusion now handles nested SliceEncodingAttr layouts
• Reading slice-encoded tiles in the unrolled path properly reconstructs a full-rank tensor for sliced layouts. The sliced dimension is padded to 1 but the padding is removed in the layout register space. This results in an appropriately ranked tensor, but the value for the padded dimension is ignored - exactly what we want when loading a sliced tile from a sliced global tensor.

I also enabled all reduction tests in GitHub Actions. I used parallel jobs but also bumped the timeout from 60m to 90m, as the addition reduction tests will add a few mins and we were pretty close to 60 to begin with.