Co-tile pre-spatial same-resolution skip connections in 2D tiling

[asteinh]

Summary

Relaxes the 2D (height and width) tiling eligibility so a stage whose
Concat/Add/Mul is consumed BEFORE its single spatial op, with
same-resolution stage-external operands (a skip connection), is admitted for 2D
tiling instead of being blanket-excluded. This lets an encoder-decoder block of
the form [Concat(up, skip), Conv] 2D-tile bit-exact against ONNX Runtime, where
previously any binary/Concat stage fell back to the 1D path or failed closed.

Compiler-only: no runtime change is required. The executor already loads a
pre-spatial same-resolution operand at the correct halo rectangle, and the
Concat/Add/Mul kernels already consume the tile geometry - the cross-repo
differential gate proves the tiled result is bit-exact.

What changed

• _stage_2d_eligible (analysis/partition_spatial.py): the blanket
  Add/Mul/Concat exclusion is replaced by a co-tileability check. A stage
  is admitted only when it has one spatial op, every Concat/Add/Mul is
  consumed strictly before that spatial op (pre-spatial), and every stage-external
  operand is a rank-4 tensor at the same H/W as the op output. Post-spatial
  operands, different-resolution operands, and constant operands stay fail-closed.
• slow_pool_usage (analysis/validation.py): the slow/PSRAM budget check now
  computes a per-tiled-stage interval-overlap concurrent-liveness peak. A boundary
  tensor is counted for a tiled stage when its lifetime overlaps the stage's
  op-step interval, so a stage's own inputs and outputs are counted concurrently
  and a long-lived skip resident across a stage is counted too. This replaces the
  per-stage input+output max, which under-counted a skip spanning multiple stages
  (and, if left as a naive per-op sample, would under-count a multi-op stage).
• Cross-repo differential gate (scripts/crossrepo_contract.py): new co-tiled
  Concat-skip (float and int8) and Add-skip 2D-tiled cases assert the emitted
  plan is TILE_AXIS_HW with both axes split and match ONNX Runtime bit-exact.

Testing

• pytest full suite: 335 passed, 2 skipped.
• Cross-repo gate: the new float_cotiled_concat_2d, int8_cotiled_concat_2d,
  and float_cotiled_add_2d cases reach TILE_AXIS_HW (tiles on both axes) and
  match ONNX Runtime (float allclose 1e-5; int8 at the established 1-LSB QDQ
  tolerance); all existing cases pass.
• New unit tests cover the admitted pre-spatial Concat/Add skip, the fail-closed
  controls (post-spatial, different-resolution, constant operand), and the
  concurrent-liveness slow-pool peak for a long-lived skip and a multi-op stage.