fix: fail closed on post-spatial external skip in 1D-height tiling to match the 2D path

[asteinh]

Summary

A rank-4 stage whose post-spatial Add/Mul/Concat consumes a stage-external,
output-resolution operand (e.g. [Conv(stride 2), Add(conv_out, skip)]) was admitted by
the 1D-height tiler, even though the 2D tiler already rejects that exact shape. The
1D-height executor loads every stage input at the spatial op's input-halo rows, so a
strided interior tile would read the external skip at stride*out_start rows -- the wrong
rows, and past the operand's height.

This shape is unreachable through the standard compile pipeline today (the greedy temporal
partition only emits a single forced op as an over-budget/tiled stage, so a multi-op stage
is never tiled), so it is a latent-correctness and path-consistency gap rather than an
active bug. It was confirmed by an adversarially-verified investigation. The fix makes the
1D-height and 2D paths reject in lockstep.

Change

• New _has_post_spatial_binary(stage_ops): true when a Concat/Add/Mul is consumed
  at or after the stage's spatial op. Shared by _stage_2d_eligible (a behavior-preserving
  refactor of its existing post-spatial check) and _stage_tile_axis.
• _stage_tile_axis returns TILE_AXIS_NONE for a rank-4 stage with a post-spatial
  binary/Concat, so it runs untiled via exec_stage_normal.
• Tests: the new post-spatial shape is rejected on both the 1D and 2D paths; a pre-spatial
  co-tiled skip stays admitted on both.

Testing

Full suite: 338 passed, 2 skipped.