docs: record sdpa backend import binding

[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；补平台绑定文档与导入回归）
• 三层测试绿: test/{kernel,op,framework}_level.py（文档/导入测试无数值语义变化；P800 三层已在 fix: guard P800 sdpa perf loading #22 复验）
• 性能对照: script/bench_perf.py（native 参照）+ reports/ 更新（无性能路径变化）
• 黄金绿: goldendata/ + script/check_accuracy.py（无数值语义变化）
• 包化回归绿: ops/test_packaging.py 思路复核（16 目录包态导入扫描通过）
• 多平台变更时: PLATFORM.md / MERGE.md 已同步（本 PR 补 PLATFORM.md）
• 署名一致: wt-YYYY-MM-DD-fix + # wt <邮箱>，commit 作者相同（小范围文档/测试跟进）

说明

Follow-up to #22, covering the two non-blocking review suggestions that can be completed without MLU hardware:

1. Document the multi-operator/perf-registry reference shadowing constraint in ops/sdpa/PLATFORM.md.
2. Explain why the import test reloads an already imported backend.
3. Extend the same foreign-top-level-reference regression to both P800 and MLU backends.

The MLU change is an import-path regression only; full ops/sdpa/example.py mlu590 still needs an MLU machine.

验证

pytest -q tests/unit/test_sdpa_backend_import.py tests/unit/test_perf.py  7 passed
pytest -q tests/unit                                                    24 passed
ops package import scan                                                  16/16
Markdown link check                                                      0 errors
intake validation                                                        0 errors
git diff --check                                                         PASS