cuda: declare the qwen3_5 backend

[xiaoyu-xyz]

Purpose

#19 added the kernels under src/backends/cuda/qwen3_5/ without declaring them, so the directory has kernel sources and no manifest. #25 defines the manifest and its checker, and against main the checker reports exactly that:

[error] qwen3_5: has kernel sources (attention.cu, common.cuh, elementwise.cu)
        but no qwen3_5.backend.json manifest

Rebased onto current main.

The change has three parts:

1. The manifest. Values come from the merged files rather than from me: build.sh writes libqwen3_5_cuda.so and defaults arch to CUDA_COMPUTE_CAP-or-89 in a single -gencode flag, and the directory README says the tensor-core kernels need sm_80 or newer.
2. The ABI version. The first commit said abi_version: 1; ops.h defines CS1_ABI_VERSION 4. Nothing caught that until cuda: add the backend contract and its checker #25's checker learned to read the header macro, which is what surfaced it.
3. The CI step. ci.yml gains a cuda-contract job: the checker's own tests, the repository's CUDA tests under tests/cuda, and check_contract.py --repo-root .. No CUDA toolkit, no GPU.

status is experimental, not validated. #19 validated the kernels, but a validated manifest asserts a tolerance and a reference entrypoint, and those are twu3202's to state rather than mine. Promoting it once a tolerance is recorded is a one-line change.

Merge order

This depends on #25. The checker exits 1 on a tree with kernels and no declaration, which is main's state without this manifest, so the job cannot land before it. The checker's own tests and check_contract.py also arrive in #25, so those two steps skip until it lands — the job comments say so, and the notice names #25 rather than reporting a bare missing file. The tests/cuda step is main's own tree and runs in either order.

Test Result

Both merge orders, run from the commit rather than from a working tree:

• cuda: declare the qwen3_5 backend #77 alone — the checker steps skip with the notice; tests/cuda runs, 8 tests, OK.
• cuda: add the backend contract and its checker #25 + cuda: declare the qwen3_5 backend #77 — 68 checker tests, 8 repository tests, and check_contract exits 0 with 2 backend(s), 0 error(s), 1 warning(s). The warning is Laya's: it declares abi_version: 1 and none of its declared sources defines an ABI_VERSION macro for the checker to compare against.

The manifest itself: it parses, all nine declared sources exist, and the directory holds no .cu, .cuh or .h file the manifest omits — that last part by hand, since the checker does not compare a directory against sources. The ABI check cross-checks ops.h's CS1_ABI_VERSION 4 against abi_version: 4.

Deliberately not included: numerics and reference. Both are about a parity claim this does not make.

[Levius-Fubuki]

Checked d1a0a12 with #25: 66 checker tests pass; current-main manifests yield zero errors and one disclosed Laya ABI warning. One CI nit: the new kernel-test loop scans beside kernels and misses root tests/cuda (8 CPU tests). Please use the repository test layout. No CUDA compilation or GPU parity was rerun.

[xiaoyu-xyz]

Fixed in b1a886a, rebased onto current main. The loop scanned src/backends/cuda/*/, so it matched the checker's own tests/ directory — running those 66 tests a second time — and never reached tests/cuda. It now uses the repository layout, the same invocation as the benchmarks job: python -m unittest discover -s tests/cuda -p 'test_*.py' -v.

tests/cuda is already in main, so that step runs in either merge order; only the checker steps stay behind the present gate. Verified from the commit in both orders, and CI on b1a886a ran the 8 tests.

Related, in #25 (184de43): a flat backend was taken to own the whole src/backends/cuda/ directory, so laya.backend.json exempted qwen3_5/ from the undeclared-backend check. Current main now reports that error, which is the state this PR clears.