Fix the bugs about operator registration by PyTorch Dispatcher (#2786)

**Background:**

There are two principles about operator registration in PyTorch
- The same namespace can be only registered once by `TORCH_LIBRARY`
- The operator signatures can be only registered once by `def`

Considering that all custom operators defined in the current repo are
only used by Ascend, instead of defining a common operator schema by
vLLM, all accelerators then follow this operator schema and complete the
implementation based on their respective hardware, which is conducive to
functional abstraction.

Therefore, we can rename the operator registration namespace to an
Ascend-specific namespace(**_C_ascend**).

Related ISSUE: https://github.com/vllm-project/vllm-ascend/issues/2742


- vLLM version: main
- vLLM main:
f592b3174b

Signed-off-by: FFFrog <ljw1101.vip@gmail.com>

tests/e2e/singlecard/ops/test_bgmv_expand.py

@@ -33,8 +33,8 @@ def test_bgmv_expand():
     y_npu = y.npu()
 
     y_out = bgmv_expand_cpu_impl(x, w, indices, y, 0, 128)
-    y_out_npu = torch.ops._C.bgmv_expand(x_npu, w_npu, indices_npu, y_npu, 0,
-                                         128)
+    y_out_npu = torch.ops._C_ascend.bgmv_expand(x_npu, w_npu, indices_npu,
+                                                y_npu, 0, 128)
 
     # Compare the results.
     torch.testing.assert_close(y_out_npu.cpu(),

tests/e2e/singlecard/ops/test_bgmv_shrink.py

@@ -33,7 +33,7 @@ def test_bgmv_shrink():
     y_npu = y.npu()
 
     y = bgmv_shrink_cpu_impl(x, w, indices, y, 0.5)
-    torch.ops._C.bgmv_shrink(x_npu, w_npu, indices_npu, y_npu, 0.5)
+    torch.ops._C_ascend.bgmv_shrink(x_npu, w_npu, indices_npu, y_npu, 0.5)
 
     # Compare the results.
     torch.testing.assert_close(y_npu.cpu(),

tests/e2e/singlecard/ops/test_rotary_embedding.py

@@ -182,7 +182,7 @@ def test_rotary_embedding_quant_with_leading_dim(
     )
 
     ref_query, ref_key = rope.forward_native(positions, query, key)
-    query, key = torch.ops._C.rotary_embedding(
+    query, key = torch.ops._C_ascend.rotary_embedding(
         positions,
         query,
         key,
@@ -239,7 +239,7 @@ class ModelwithRotaryEmbedding(nn.Module):
         # we simulated a simple attention layer to test if it can be seamlessly captured into aclgraph
         qkv = self.qkv_proj(hidden_states)
         q, k, v = qkv.chunk(3, dim=-1)
-        query, key = torch.ops._C.rotary_embedding(
+        query, key = torch.ops._C_ascend.rotary_embedding(
             positions,
             q,
             k,
@@ -299,7 +299,7 @@ def test_capture_rotary_embedding_in_aclgraph(
         # Validate if the rotary_embedding custom kernel is indeed inside the graph by
         # string match
         graph = str(gm.graph)
-        assert "_C.rotary_embedding" in graph
+        assert "_C_ascend.rotary_embedding" in graph
         return gm
 
     static_positions = torch.randint(0, max_position_embeddings,

tests/e2e/singlecard/ops/test_vocabparallelembedding.py

@@ -72,7 +72,7 @@ def test_get_masked_input_and_mask(
 
     # Get custom op result
     print("input_tensor:", input_tensor)
-    custom_masked_input, custom_mask = torch.ops._C.get_masked_input_and_mask(
+    custom_masked_input, custom_mask = torch.ops._C_ascend.get_masked_input_and_mask(
         input_tensor, test_case["org_start"], test_case["org_end"],
         test_case["padding"], test_case["added_start"], test_case["added_end"])