build(ascend-custom): pip-buildable custom kernels via pre-built static lib

`BUILD_CUSTOM_KERNEL=ON` under `scikit-build-core` / `pip install` trips a
`CANN` `extract_host_stub.py` path-handling bug (`KeyError` on the
`/./workspace/...` path produced by `$<TARGET_OBJECTS>` in CMake 4.x).
`CMakeLists.txt` already notes the issue and suggests the standalone
`src/ascend/custom/build.sh` script as the workaround.

- `src/CMakeLists.txt`: when a pre-built
  `src/ascend/custom/build/lib/libno_workspace_kernel.a` exists, import
  it as `IMPORTED STATIC` and link via `$<TARGET_FILE:…>` +
  `--whole-archive`.  Falls back to in-tree `ascendc_library()` when the
  pre-built is absent (works for direct `cmake` builds outside `pip`).
- `src/ascend/custom/CMakeLists.txt`: pushes the main `src/` onto the
  interface target's `INCLUDES` property so the kernel TU can
  `#include "data_type.h"` — the `ascendc_library()` macro forwards this
  into its nested `ExternalProject_Add` sub-build.
- `src/ascend/custom/add_rms_norm/op_kernel/.clang-tidy`: disables all
  `clang-tidy` checks for device code under this directory (compiled by
  `ccec`, absent from `compile_commands.json`, `__aicore__` macro parses
  incorrectly without `kernel_operator.h`).

Author: zhangyue

src/CMakeLists.txt

@@ -242,7 +242,25 @@ if(WITH_ASCEND)
 
     # Custom `AscendC` kernels (PyTorch extension, requires `torch_npu`).
     if(BUILD_CUSTOM_KERNEL)
-        add_subdirectory(ascend/custom)
+        # `scikit-build-core`'s temp-dir builds trip the `CANN`
+        # `extract_host_stub.py` path-handling bug (`KeyError` on
+        # `/./workspace/...` paths in `$<TARGET_OBJECTS>`).  Workaround:
+        # prefer a pre-built static lib produced by the standalone
+        # `src/ascend/custom/build.sh` script.  Fall back to in-tree
+        # `ascendc_library()` when no pre-built is available (works for
+        # direct `cmake` builds outside `pip`).
+        set(_prebuilt_custom_lib
+            "${CMAKE_CURRENT_SOURCE_DIR}/ascend/custom/build/lib/libno_workspace_kernel.a")
+
+        if(EXISTS "${_prebuilt_custom_lib}")
+            message(STATUS "Custom `AscendC` kernels: using pre-built `${_prebuilt_custom_lib}`.")
+            add_library(no_workspace_kernel STATIC IMPORTED GLOBAL)
+            set_target_properties(no_workspace_kernel PROPERTIES
+                IMPORTED_LOCATION "${_prebuilt_custom_lib}")
+        else()
+            message(STATUS "Custom `AscendC` kernels: building in-tree (run `src/ascend/custom/build.sh` first to bypass the `CANN` path bug under `pip`).")
+            add_subdirectory(ascend/custom)
+        endif()
 
         # Link the compiled `AscendC` kernel objects into `infiniops` so that
         # custom kernel implementations (e.g. `RmsNorm` index 1) can call
@@ -379,9 +397,11 @@ if(GENERATE_PYTHON_BINDINGS)
     # The `Operator<..., 1>` template instantiations that call
     # `aclrtlaunch_*` live in `ops.cc`, so link here with
     # `--whole-archive` to ensure all launch functions are available.
+    # `$<TARGET_FILE>` works for both real `ascendc_library()` targets and
+    # `IMPORTED` targets pointing at a pre-built `.a`.
     if(BUILD_CUSTOM_KERNEL)
         target_link_libraries(ops PRIVATE
-            -Wl,--whole-archive no_workspace_kernel -Wl,--no-whole-archive)
+            -Wl,--whole-archive $<TARGET_FILE:no_workspace_kernel> -Wl,--no-whole-archive)
     endif()
 
     set_target_properties(infiniops PROPERTIES INSTALL_RPATH "$ORIGIN")

src/ascend/custom/CMakeLists.txt

@@ -63,6 +63,15 @@ ascendc_library(no_workspace_kernel STATIC
     ${PROJECT_OP_SRC_BASE}/add_rms_norm/op_kernel/add_rms_norm.cpp
 )
 
+# The kernel translation units include `"data_type_enum.h"` from the main
+# project's `src/` so that launcher and device code share one `DataType`
+# enum.  `ascendc_library` forwards the interface target's `INCLUDES`
+# property to the nested `ExternalProject_Add` (see
+# `${ASCEND_HOME_PATH}/tools/tikcpp/ascendc_kernel_cmake/legacy_modules/function.cmake`),
+# so append the main `src/` dir here.
+set_property(TARGET no_workspace_kernel_interface APPEND PROPERTY
+    INCLUDES ${PROJECT_OP_SRC_BASE}/../..)
+
 # Create the shared library `libascend_kernel.so`.
 add_library(${OP_PLUGIN_NAME} SHARED ${OP_SRCS})
 

src/ascend/custom/add_rms_norm/op_kernel/.clang-tidy

@@ -0,0 +1,9 @@
+---
+# `op_kernel/*.cpp` is `AscendC` device code compiled by `ccec`, not by
+# the host toolchain, so it has no entry in `compile_commands.json` and
+# `clang-tidy` cannot parse it correctly (the `__aicore__` macro expands
+# unexpectedly when `kernel_operator.h` is absent).  Disable all checks
+# here — the `op_host/` side and the `kernel_custom.h` launcher still
+# enforce the full ruleset.
+
+Checks: '-*'