feat: SGLang-style paged attention kernels replace page-table path

- PagedAttentionParams uses flat KV pool + req_to_token + kv_indptr/qo_indptr instead of page_table
- MMA split-KV decode and split-Q prefill kernels with indirect ragged-batch addressing
- Prefill kernel accepts 4D mask (causal-aware); decode kernel supports 2D mask
- CudaBackend is inference-only: kv_cache=None raises, no torch fallback
- benchmark.py: required --ckpt, --backend/--compare options
- Parallel build isolates build-temp/build-lib per subprocess
- Standalone test covers decode/prefill with mask, 27 cases pass

csrc/kernels/attn_paged_decode.cu

@@ -3,23 +3,23 @@
 
 torch::Tensor attn_paged_decode(
     torch::Tensor q,
-    torch::Tensor page_table,
     torch::Tensor k_cache,
     torch::Tensor v_cache,
-    int64_t page_size,
-    int64_t kv_len,
+    torch::Tensor req_to_token,
+    torch::Tensor req_pool_indices,
+    torch::Tensor kv_indptr,
+    int64_t max_seq_len,
     c10::optional<torch::Tensor> mask,
     int64_t causal_offset,
-    double scale,
-    int64_t layout
+    double scale
 ) {
     PagedAttentionParams<bf16> p;
-    attn_pack_paged_params(q, page_table, k_cache, v_cache,
-                           page_size, kv_len, mask, causal_offset, scale, layout, p);
+    attn_pack_paged_decode_params(q, k_cache, v_cache,
+                                   req_to_token, req_pool_indices, kv_indptr,
+                                   max_seq_len, mask, causal_offset, scale, p);
 
-    auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
-    auto O_view = (layout == BLHD) ? O.transpose(1, 2) : O;
-    p.o = (bf16*)O_view.data_ptr();
+    auto O = torch::empty({q.size(0), q.size(1), q.size(2)}, q.options());
+    p.o = (bf16*)O.data_ptr();
 
     alloc_split_partials(p);
     DISPATCH_HEAD_DIM(p.head_dim, dispatch_paged_decode, p);
@@ -30,14 +30,14 @@ torch::Tensor attn_paged_decode(
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     m.def("attn_paged_decode", &attn_paged_decode,
         py::arg("q"),
-        py::arg("page_table"),
         py::arg("k_cache"),
         py::arg("v_cache"),
-        py::arg("page_size"),
-        py::arg("kv_len"),
+        py::arg("req_to_token"),
+        py::arg("req_pool_indices"),
+        py::arg("kv_indptr"),
+        py::arg("max_seq_len"),
         py::arg("mask") = py::none(),
         py::arg("causal_offset") = -1,
         py::arg("scale") = 0.0,
-        py::arg("layout") = (int64_t)BHLD,
-        "Paged GQA decode — split-KV with direct page-table access.");
+        "SGLang-style paged decode: flat KV pool + req_to_token + kv_indptr.");
 }