[codex] add test6 validshape tpush/tpop sample

[zhangstevenunity]

What changed

• add test/samples/TPushTPop/test6/kernel.pto
• base the new sample on the existing TPushTPop/test2 A5 case structure
• make the cube-side accumulator tile dynamic in valid shape
• call pto.set_validshape %acc_tile, %c8, %c16 after pto.tmatmul
• pop the result on the vector side as a 4x16 tile with split = 1

Why

This adds a dedicated sample for the flow you requested: after updating the ACC tile valid shape, pop the result from the vector kernel.

Validation

• compiled with an existing local WSL ptoas
• command used:
  • /home/rdp/hw-native-sys/PTOAS/build/tools/ptoas/ptoas --pto-arch=a5 --enable-insert-sync -o /tmp/ptoas-test6/kernel.cpp /mnt/c/Users/rdp/Documents/ptoas/test/samples/TPushTPop/test6/kernel.pto
• confirmed generated output contains:
  • SetValidShape(...)
  • Tile<TileType::Vec, float, 4, 16, ...>
  • TPOP<..., Tile<TileType::Vec, float, 4, 16, ...>, TileSplitAxis::TILE_UP_DOWN>

Notes

• the local worktree also contains unrelated edits under test/samples/TPushTPop/test2/kernel.pto; those were intentionally excluded from this PR

[gemini-code-assist]

Code Review

This pull request introduces a new test case for matrix multiplication using TPush and TPop operations between cube and vector kernels. A potential logic error was identified in the cube kernel where the accumulator tile is allocated outside the loop; because pto.set_validshape modifies tile metadata in-place, subsequent loop iterations would incorrectly perform partial multiplications. It is recommended to move the tile allocation inside the loop to ensure each iteration starts with the correct dimensions.

[gemini-code-assist]

The accumulator tile %acc_tile is allocated outside the loop, but its valid shape is modified inside the loop using pto.set_validshape (line 54). Since set_validshape updates the tile metadata in-place, subsequent iterations of the loop will start with the modified valid shape (8x16 instead of 16x16). This will likely cause the pto.tmatmul in iterations 2-4 to only compute a partial result (8x16) instead of the full 16x16 multiplication performed in the first iteration. To ensure each iteration computes a full 16x16 result before shrinking the valid shape for pushing, the alloc_tile for %acc_tile should be moved inside the loop, similar to the implementation in test/samples/TPushTPop/test4/kernel.pto.