fix(nccl): fix stream argument passing to CUDA functions

Author: dssgabriel

src/KokkosComm/nccl/req.hpp

@@ -70,11 +70,11 @@ class Req<Experimental::Nccl> {
 
   friend void wait(Req<Experimental::Nccl> &req);
   friend void wait_all(std::span<Req<Experimental::Nccl>> reqs);
-  friend int wait_any(std::span<Req<Experimental::Nccl>> reqs);
+  friend void wait_any(std::span<Req<Experimental::Nccl>> reqs);
 };
 
 inline auto wait(Req<Experimental::Nccl> &req) -> void {
-  cudaStreamSynchronize(&req.get_inner());
+  cudaStreamSynchronize(req.get_inner());
   for (auto &f : req.record_->postWaits_) {
     f();
   }
@@ -92,7 +92,7 @@ inline auto wait_any(std::span<Req<Experimental::Nccl>> reqs) -> void {
   // Loop while we don't have at least one completed request.
   while (completed == 0) {
     for (Req<Experimental::Nccl> &req : reqs) {
-      auto res = cudaStreamQuery(&req.get_inner());
+      auto res = cudaStreamQuery(req.get_inner());
 
       // If the current request has completed, we must make sure the post-wait callbacks run and are cleared.
       // Calling `wait` should be a no-op if the request has no callback to execute.