Fix crash in TileLoops when redomap input array is also used in postlude

[Copilot]

mmBlkRegTiling produces invalid IR when a SegMap contains two redomaps sharing an input array — the first redomap is tiled (its input slice goes into tab_inv_stm for shared-memory loading), but the second redomap in the postlude still references that same slice variable, which no longer exists in the epilogue's scope.

Changes

• src/Futhark/Optimise/BlkRegTiling.hs — Added a guard in matchesBlkRegTile that returns Nothing if any key in tab_inv_stm (tiled redomap inputs) is also free in the postlude code (code2'' <> code2). This causes mmBlkRegTiling to decline and fall through to tileInKernelBody, which handles the case correctly by placing invariant prelude definitions in the outer kernel body where they remain in scope.

• tests/tiling/issue2467.fut — Regression test using a matrix multiply + entrywise divide pattern that triggers the two-redomap-sharing-an-input scenario:

entry main W A WH =
  let W_TA  = matmul (transpose W) A
  let W_TWH = matmul (transpose W) WH
  in matdiv_entrywise W_TA W_TWH

Previously this crashed with Use of unknown variable as_transformed_row_... after the tile-loops pass.

[athas]

This is a pretty obvious fix, close to what I would have written myself. My only worry is whether it is too conservative, and would stop tiling in other cases. However, it seems clear to me that all cases caught by this check would previously result in invalid code.