refactor: template combine kernel, fix mask bug, unify dispatch

- Template combine kernel, share macros, extract entry_utils helpers
- Fix mask indexing (pass stride not pre-multiplied base)
- Remove !p.use_mask — MMA handles mask

csrc/kernels/attn_prefill.cu

@@ -47,7 +47,7 @@ torch::Tensor attn_prefill(
     auto O_view = (layout == 1) ? O.transpose(1, 2) : O;
     p.o = (bf16*)O_view.data_ptr();
 
-    dispatch_head_dim(p.head_dim, [&]<int D>() { dispatch_prefill<D>(p); });
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_prefill, p);
     return O;
 }