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/transpose.rs

@@ -25,7 +25,7 @@ pub fn reshape_heads_gpu(x: &Tensor, seq_len: usize, num_heads: usize, head_dim:
     unsafe {
         launch_reshape_heads_bf16(
             x.data_ptr() as _, out.data_ptr() as *mut c_void,
-            seq_len as i32, num_heads as i32, head_dim as i32, std::ptr::null_mut(),
+            seq_len as i32, num_heads as i32, head_dim as i32, xserv_cuda::current_stream_raw(),
         );
     }
     out
@@ -40,7 +40,7 @@ pub fn merge_heads_gpu(x: &Tensor, seq_len: usize, num_heads: usize, head_dim: u
     unsafe {
         launch_merge_heads_bf16(
             x.data_ptr() as _, out.data_ptr() as *mut c_void,
-            seq_len as i32, num_heads as i32, head_dim as i32, std::ptr::null_mut(),
+            seq_len as i32, num_heads as i32, head_dim as i32, xserv_cuda::current_stream_raw(),
         );
     }
     out
@@ -54,7 +54,7 @@ pub fn transpose_for_rope_gpu(x: &Tensor, seq_len: usize, num_heads: usize, head
     unsafe {
         launch_transpose_hsd_to_shd_bf16(
             x.data_ptr() as _, out.data_ptr() as *mut c_void,
-            seq_len as i32, num_heads as i32, head_dim as i32, std::ptr::null_mut(),
+            seq_len as i32, num_heads as i32, head_dim as i32, xserv_cuda::current_stream_raw(),
         );
     }
     out
@@ -68,7 +68,7 @@ pub fn transpose_from_rope_gpu(x: &Tensor, seq_len: usize, num_heads: usize, hea
     unsafe {
         launch_transpose_shd_to_hsd_bf16(
             x.data_ptr() as _, out.data_ptr() as *mut c_void,
-            seq_len as i32, num_heads as i32, head_dim as i32, std::ptr::null_mut(),
+            seq_len as i32, num_heads as i32, head_dim as i32, xserv_cuda::current_stream_raw(),
         );
     }
     out
@@ -87,7 +87,7 @@ pub fn repeat_kv_gpu(x: &Tensor, n_rep: usize) -> Tensor {
     unsafe {
         launch_repeat_kv_bf16(
             x.data_ptr() as _, out.data_ptr() as *mut c_void,
-            kv_heads as i32, n_rep as i32, seq_len as i32, head_dim as i32, std::ptr::null_mut(),
+            kv_heads as i32, n_rep as i32, seq_len as i32, head_dim as i32, xserv_cuda::current_stream_raw(),
         );
     }
     out
@@ -126,14 +126,14 @@ pub fn strided_to_contiguous_gpu(x: &Tensor) -> Tensor {
                 numel as i32, ndim as i32,
                 shape4[0], shape4[1], shape4[2], shape4[3],
                 strides4[0], strides4[1], strides4[2], strides4[3],
-                in_offset, std::ptr::null_mut(),
+                in_offset, xserv_cuda::current_stream_raw(),
             ),
             DType::F32 => launch_strided_copy_f32(
                 storage_ptr as _, out.data_ptr() as *mut c_void,
                 numel as i32, ndim as i32,
                 shape4[0], shape4[1], shape4[2], shape4[3],
                 strides4[0], strides4[1], strides4[2], strides4[3],
-                in_offset, std::ptr::null_mut(),
+                in_offset, xserv_cuda::current_stream_raw(),
             ),
             _ => panic!("strided_to_contiguous_gpu: unsupported dtype {:?}", x.dtype()),
         }