cuda: infrastructure for whole-step CUDA graph capture

- Thread-local launch stream (xserv_cuda::stream): every kernel
  wrapper, cublasSetStream, and NCCL collective now launches on
  current_stream_raw() — the legacy null stream by default (behavior
  unchanged), or the capture stream installed via push_stream during
  graph capture. Capture is impossible on the legacy stream.
- Allocator retain mode: blocks freed inside a retain window are
  quarantined (RetainedBlocks) instead of pooled, so an instantiated
  graph keeps exclusive ownership of every intermediate buffer it
  references across replays.
- Capture mode GLOBAL -> THREAD_LOCAL: concurrent TP rank threads
  must not poison each other's captures with their own cudaMallocs.
- embedding_device_ids / rope_inplace_device_pos: variants reading
  token ids / positions from persistent device buffers, replacing the
  per-call host upload that a captured region cannot contain.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

crates/xserv-kernels/src/moe.rs

@@ -100,7 +100,7 @@ pub fn moe_topk_softmax(
             topk_ids.data_ptr() as *mut c_void,
             topk_weights.data_ptr() as *mut c_void,
             num_tokens as i32, num_experts as i32, top_k as i32,
-            std::ptr::null_mut(),
+            xserv_cuda::current_stream_raw(),
         );
     }
 
@@ -121,7 +121,7 @@ pub fn moe_replicate(x: &Tensor, local_experts: usize) -> Tensor {
             x.data_ptr() as *const c_void,
             out.data_ptr() as *mut c_void,
             num_tokens as i32, hidden as i32, local_experts as i32,
-            std::ptr::null_mut(),
+            xserv_cuda::current_stream_raw(),
         );
     }
 
@@ -144,7 +144,7 @@ pub fn moe_bias_add_3d(x: &Tensor, bias: &Tensor) {
             x.data_ptr() as *mut c_void,
             bias.data_ptr() as *const c_void,
             batch as i32, num_tokens as i32, dim as i32,
-            std::ptr::null_mut(),
+            xserv_cuda::current_stream_raw(),
         );
     }
 }
@@ -177,7 +177,7 @@ pub fn moe_weighted_sum(
             out.data_ptr() as *mut c_void,
             num_tokens as i32, hidden as i32, top_k as i32,
             expert_start as i32, local_experts as i32,
-            std::ptr::null_mut(),
+            xserv_cuda::current_stream_raw(),
         );
     }
 
@@ -224,7 +224,7 @@ pub fn moe_sparse_gemv_fp8(
             y.data_ptr() as *mut c_void,
             num_tokens as i32, n as i32, k as i32, top_k as i32,
             expert_start as i32, local_experts as i32, x_per_slot as i32,
-            std::ptr::null_mut(),
+            xserv_cuda::current_stream_raw(),
         );
     }
     y
@@ -256,7 +256,7 @@ pub fn moe_sparse_gemv_mxfp4(
             y.data_ptr() as *mut c_void,
             num_tokens as i32, n as i32, k as i32, top_k as i32,
             expert_start as i32, local_experts as i32, x_per_slot as i32,
-            std::ptr::null_mut(),
+            xserv_cuda::current_stream_raw(),
         );
     }
     y
@@ -288,7 +288,7 @@ pub fn moe_weighted_sum_sparse(
             out.data_ptr() as *mut c_void,
             num_tokens as i32, hidden as i32, top_k as i32,
             expert_start as i32, local_experts as i32,
-            std::ptr::null_mut(),
+            xserv_cuda::current_stream_raw(),
         );
     }
     out
@@ -338,7 +338,7 @@ pub fn batched_gemm_strided(a: &Tensor, b: &Tensor) -> Tensor {
 
     let handle = cublas_handle();
     unsafe {
-        cublasSetStream_v2(handle, std::ptr::null_mut());
+        cublasSetStream_v2(handle, xserv_cuda::current_stream_raw());
         let status = cublasGemmStridedBatchedEx(
             handle,
             0, 0, // CUBLAS_OP_N, CUBLAS_OP_N