[Test] Consolidate test cleanup and CI taxonomy (net -11.4K lines) (#37436)

Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>

python/sglang/kernels/ops/diffusion/README.md

@@ -17,7 +17,7 @@ from sglang.kernels.ops.diffusion import fused_rmsnorm_scale_shift_bitexact
 ```
 
 **Import from the package, never from a submodule.** The internal layout is
-free to move; the facade is not. `test_import_surface.py` enforces this, with
+free to move; the facade is not. Callers should use the facade, with
 a small allowlist for tests that deliberately exercise one backend.
 
 Resolution is lazy (PEP 562): the backends have disjoint heavy dependencies
@@ -189,7 +189,7 @@ inspecting model modules is its whole job.
    generated by the KDA workflow in `sglang.kernels.kda_kernels`, together
    with its source revision and any JIT CUDA source files.
 2. Export it from `__init__.py` (`_EXPORTS`) and register a `KernelSpec`
-   (`_SPECS`) — `test_import_surface.py` checks both resolve.
+   (`_SPECS`).
 3. Give it a `can_use_*` predicate; raise, don't return `None`.
 4. State the numerical contract in the module docstring, including which
    shapes it was verified on.

python/sglang/kernels/ops/diffusion/__init__.py

@@ -5,8 +5,8 @@ This module is the **only** supported import surface for these kernels::
     from sglang.kernels.ops.diffusion import fused_rmsnorm_scale_shift_bitexact
 
 Importing a submodule directly (``...diffusion.norm.norm_triton``) couples the
-caller to the file layout; ``test_import_surface.py`` guards against it.  The
-one exception is a test that deliberately exercises a single backend.
+caller to the file layout. The one exception is a test that deliberately
+exercises a single backend.
 
 Layout -- ordinary implementations use one subpackage per **operator domain**
 (``norm``, ``modulate``, ``rope``, ``activation``, ``attention``, ``routing``,