fix: upgrade legacy operator register contracts

[TruNcat3]

自查

• 核心八件齐全: README / REPORT / __init__.py（包化入口） / reference / register / example / kernel/ / test/（不新增算子；只升级既有 register）
• 三层测试绿: test/{kernel,op,framework}_level.py（本环境无 Cambricon MLU；新增 CPU 单元契约测试 + 包化/脚本态导入回归）
• 性能对照: script/bench_perf.py（native 参照）+ reports/ 更新（注册签名修复无 kernel 语义变化，不重采性能）
• 黄金绿: goldendata/ + script/check_accuracy.py（无算子语义变化）
• 包化回归绿: python3 ops/test_packaging.py（全算子导入扫描）
• 多平台变更时: PLATFORM.md / MERGE.md 已同步（仍为单平台；chore: PR#5 的 13 个算子按新库约定补两项——register 签名升级 + 多平台文档（平台扩展前） #19 明确第二平台前可欠）
• 署名一致: wt-YYYY-MM-DD-fix + # wt <邮箱>，commit 作者相同（本 PR 为库级修复，无新算法文件署名块）

说明

Fixes #19.

register contract

All 13 legacy Cambricon operators now expose:

register_a1(
    dispatch_key="AutogradPrivateUse1",
    counter: dict | None = None,
    platform: str | None = None,
)

• optional counter["n"] increments at every intercepted aten call
• explicit platform="cambricon" / "mlu" / "mlu590" is validated
• supported dispatch keys are validated before creating a torch library
• detach_, empty, and result_type intentionally remain non-registerable while accepting the new contract
• kernel imports now use package-relative imports first and preserve standalone single-operator script mode fallback

CI repair

The newly added ops-packaging job is retained, and the truncated docs-and-syntax script is restored. Markdown links, Python syntax, YAML, intake manifest, and perf-baseline structure checks are active again.

验证

pytest tests/unit -q                 24 passed
ops package import scan              16/16
standalone register import           13/13
Markdown link check                  0 errors
Python syntax                        0 errors
YAML / intake / perf baseline        PASS
git diff --check                     PASS

No Cambricon MLU device is present in this execution environment, so hardware kernel/op/framework reruns were not possible. The change is registration-only and adds CPU-visible contract tests that fake the torch library to verify interception counters without globally registering aten ops.

[TruNcat3]

误处理了：#19 属于 PR #5 的 13 个 Cambricon 算子维护者，不属于我们当前负责的三个计算算子范围。关闭此 PR，避免混淆 issue 范围。