fix(EX): corex ivcore10 build flags + deploy pipeline + topk kernel cleanup

Real machine log (2d5232c dockerrizhi.txt) shows two AST call chain breaks:

1. EVERY layer EVERY token:
   _custom_ops.py:58 'ixformer.functions has no attribute vllm_moe_topk_softmax'
   -> FusedMoE falls to PyTorch loop (2304 calls/token)

2. EVERY GDN layer (4 layers):
   'NaN in prefill GatedDeltaNet layer N (frac=0.9998-1.0000)'
   -> _torch_chunk_gated_delta_rule produces all-NaN

Fixes:
- build.sh: --cuda-gpu-arch=ivcore10, -D__ILUVATAR__ flags from real log
- Dockerfile: add ex_engine build before patch_ops
- patch_ops.sh: deploy .so + python into vllm model dir
- ex_loader.py: search co-located .so paths
- patch_model.py: remove premature auto-apply
- factor_moe_topk_softmax.cu: remove dead parallel branch

ex_engine/python/ex_loader.py

@@ -156,13 +156,27 @@ class EXEngine:
             return False
 
     def load_all(self) -> int:
-        """Load all available factor .so files from build_dir."""
+        """Load all available factor .so files from build_dir or co-located."""
         loaded = 0
+        # Search paths: build_dir first, then directory containing this module
+        search_dirs = [self.build_dir]
+        module_dir = os.path.dirname(os.path.abspath(__file__))
+        if module_dir not in search_dirs:
+            search_dirs.append(module_dir)
+        # Also check parent's build dir
+        parent_build = os.path.join(os.path.dirname(module_dir), "build")
+        if parent_build not in search_dirs:
+            search_dirs.append(parent_build)
+
         for fid in range(EX_FACTOR_COUNT):
-            so_path = os.path.join(self.build_dir, f"ex_factor_{fid}.so")
-            if self.load_factor(fid, so_path):
-                loaded += 1
-        logger.info("EX Engine: loaded %d/%d factors", loaded, EX_FACTOR_COUNT)
+            for d in search_dirs:
+                so_path = os.path.join(d, f"ex_factor_{fid}.so")
+                if os.path.exists(so_path):
+                    if self.load_factor(fid, so_path):
+                        loaded += 1
+                    break
+        logger.info("EX Engine: loaded %d/%d factors from %s", loaded, EX_FACTOR_COUNT,
+                    search_dirs)
         return loaded
 
     def has_factor(self, factor_id: int) -> bool:

ex_engine/python/patch_model.py

@@ -191,11 +191,7 @@ def _patch_gdn_prefill(engine):
 
 
 # ---------------------------------------------------------------------------
-# Auto-apply on import if build dir exists
+# Call apply_patches() explicitly AFTER vllm model modules are loaded.
+# Integration point: qwen3_5.py calls this at the end of model __init__,
+# or patch_ops.sh adds it to the startup sequence.
 # ---------------------------------------------------------------------------
-_AUTO_BUILD_DIR = os.environ.get("EX_ENGINE_BUILD_DIR", "/workspace/ex_engine/build")
-if os.path.isdir(_AUTO_BUILD_DIR):
-    try:
-        apply_patches(_AUTO_BUILD_DIR)
-    except Exception as e:
-        logger.warning("EX Engine auto-apply failed: %s", e)