fix: extern declarations for gather_swa functions in gather_kv.cu

dsv4/kernels/cuda/gather_kv.cu

@@ -101,6 +101,17 @@ void gather_kv_cuda(
 }
 
 
+// Forward declarations for gather_swa.cu functions
+// (both files compiled together into one .so)
+extern void gather_swa_cuda(
+    torch::Tensor swa_fp8, torch::Tensor swa_rope, torch::Tensor swa_inv,
+    torch::Tensor swa_pos, torch::Tensor request_slots, torch::Tensor output,
+    int64_t head_dim, int64_t rope_dim);
+extern void gather_all_compressed_cuda(
+    torch::Tensor entries_fp8, torch::Tensor entries_rope, torch::Tensor inv_scale,
+    torch::Tensor block_table, torch::Tensor block_lens, torch::Tensor output,
+    int64_t entries_per_block, int64_t rope_dim);
+
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     m.def("gather_kv", &gather_kv_cuda, "Gather KV entries into dense tile");
     // gather_swa + gather_all_compressed are defined in gather_swa.cu