fix(cute_dsl/attention): use string literals for fence_proxy (cu13 wheel compat)

Fixes #3071.

`cute.arch.fence_proxy` is documented as taking string literals
(``"async.shared"``, ``"cta"``); the wrapper does ``ProxyKind.from_str`` /
``SharedSpace.from_str`` internally.  The enum re-exports
``cute.arch.ProxyKind`` and ``cute.arch.SharedSpace`` are gated upstream
behind ``cutlass_dsl.target_version(exact_version="12.9")`` and so are
absent from the cu13 wheel that the ``flashinfer-ci-cu130`` docker image
uses, producing:

  AttributeError: module 'cutlass.cute.arch' has no attribute 'ProxyKind'

at correction.py:267 and softmax.py:448.  On cu12 wheels the enum form
still works but emits ``DeprecationWarning: Passing enum member directly
to SharedSpace.from_str() is deprecated. Please use string literals
instead`` (visible in our local test runs).

Switching the two call sites to the string form fixes both the cu13
breakage and the cu12 deprecation warning, and is the documented stable
API on every flashinfer-supported cutlass-dsl 4.4.x wheel.

Verified: python -m pytest tests/attention/test_modular_fmha_prefill.py \
                   tests/attention/test_cute_dsl_mla_decode.py
  → 456 passed, 20 skipped (no SharedSpace/ProxyKind warnings emitted).
Made-with: Cursor

Author: pgera

flashinfer/cute_dsl/attention/roles/correction.py

@@ -263,10 +263,7 @@ def epilog(
             cute.copy(tiled_smem_store, tSMrO, tTMEM_LOADsO_i)
 
         # fence view async shared
-        cute.arch.fence_proxy(
-            cute.arch.ProxyKind.async_shared,
-            space=cute.arch.SharedSpace.shared_cta,
-        )
+        cute.arch.fence_proxy("async.shared", space="cta")
 
     @cute.jit
     def run(

flashinfer/cute_dsl/attention/roles/softmax.py

@@ -444,10 +444,7 @@ def softmax_epilog(
             tSMrO.store(o_vec.to(self.o_dtype))
             cute.copy(tiled_smem_store, tSMrO, tTMEM_LOADsO_i)
 
-        cute.arch.fence_proxy(
-            cute.arch.ProxyKind.async_shared,
-            space=cute.arch.SharedSpace.shared_cta,
-        )
+        cute.arch.fence_proxy("async.shared", space="cta")
 
     # For both softmax0 and softmax1 warp group
     @cute.jit