refactor(attn): init hisparse_coordinator before attn_backend; replace lazy property with init-time capture (#26012)

Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

test/manual/attention/test_flashattn_backend.py

@@ -91,6 +91,7 @@ class MockModelRunner:
             device=self.device,
             enable_memory_saver=False,
         )
+        self.hisparse_coordinator = None
 
 
 @unittest.skipIf(not torch.cuda.is_available(), "Test requires CUDA")

test/manual/attention/test_flashattn_mla_backend.py

@@ -76,6 +76,7 @@ class MockModelRunner:
             device=self.device,
             enable_memory_saver=False,
         )
+        self.hisparse_coordinator = None
 
 
 class MockReqToTokenPool:

test/manual/attention/test_trtllm_mla_backend.py

@@ -268,6 +268,7 @@ class MockModelRunner:
             device=self.device,
             enable_memory_saver=False,
         )
+        self.hisparse_coordinator = None
 
 
 def compare_outputs(trtllm_out, reference_out, tolerance=1e-2):

test/registered/kernels/test_dsa_indexer.py

@@ -252,6 +252,7 @@ class MockModelRunner:
                 "dsa_decode_backend": "fa3",
             },
         )()
+        self.hisparse_coordinator = None
 
 
 @unittest.skipIf(not torch.cuda.is_available(), "Test requires CUDA")