[Lane 1 · numerics] TransposedConv2d — PyTorch reference cases for conv_transpose2d #1372

Sub-issue of #1371 ([Feature]: TransposedConv2d — convTranspose2d op, CPU kernel, backward and StableHLO lowering).

Lane: 1 · Numerics/Research
Skill needed: PyTorch. No Kotlin and no SKaiNET knowledge needed.
Size: s (a few hours)
Blocked by: nothing

What to do

1. Write a small PyTorch script calling torch.nn.functional.conv_transpose2d on seeded random inputs for: stride (1,1)/(2,2)/(2,1); padding (0,0)/(1,1); outputPadding (0,0)/(1,1) (with stride 2); dilation (1,1)/(2,2); groups 1/2; with and without bias. Keep tensors tiny (e.g. N=1, C_in=4, 5×5, kernel 3×3).
2. Add the cases to the PyTorch ground-truth suite used by skainet-test/skainet-test-groundtruth (see skainet-test/skainet-test-groundtruth/TESTING.md), or, if that is too much setup, attach the script plus its JSON output (inputs, weight, bias, expected output, shapes) to this issue.
3. Note any PyTorch behaviour that surprised you (e.g. which combinations raise) in a comment.

Acceptance

• Script and expected outputs are in the repo or attached here
• Every parameter in the list above is covered at least once
• Result reported back on the parent issue

Notes

The output-size formula is in TransposedConv2d.outputSize(); your cases double as a check of it.

Coverage matrix: docs/modules/ROOT/pages/reference/ops-status-matrix.adoc · How lanes work: Issue taxonomy