Use one cusparse handle per thread to avoid race condition on cuspars… (#544)

## Summary by CodeRabbit

* **Refactor**
  * Improved GPU pointer and stream configuration to make GPU-accelerated solvers more consistent and reliable.
* **Tests**
  * Re-enabled a previously-skipped barrier solver test so it now runs as part of test suite.
* **Chores**
  * Restored default threading behavior in CI test runs by removing a forced thread-count override.

Authors:
  - Hugo Linsenmaier (https://github.com/hlinsen)
  - Ramakrishnap (https://github.com/rgsl888prabhu)

Approvers:
  - Rajesh Gandham (https://github.com/rg20)

URL: #544

Author: hlinsen

cpp/src/dual_simplex/cusparse_view.cu

@@ -138,6 +138,10 @@ cusparse_view_t<i_t, f_t>::cusparse_view_t(raft::handle_t const* handle_ptr,
     d_minus_one_(f_t(-1), handle_ptr->get_stream()),
     d_zero_(f_t(0), handle_ptr->get_stream())
 {
+  RAFT_CUBLAS_TRY(raft::linalg::detail::cublassetpointermode(
+    handle_ptr->get_cublas_handle(), CUBLAS_POINTER_MODE_DEVICE, handle_ptr->get_stream()));
+  RAFT_CUSPARSE_TRY(raft::sparse::detail::cusparsesetpointermode(
+    handle_ptr->get_cusparse_handle(), CUSPARSE_POINTER_MODE_DEVICE, handle_ptr->get_stream()));
   // TMP matrix data should already be on the GPU
   constexpr bool debug = false;
   if (debug) { printf("A hash: %zu\n", A.hash()); }

cpp/src/linear_programming/solve.cu

@@ -672,9 +672,8 @@ optimization_problem_solution_t<i_t, f_t> run_concurrent(
   // Initialize the dual simplex structures before we run PDLP.
   // Otherwise, CUDA API calls to the problem stream may occur in both threads and throw graph
   // capture off
-  auto barrier_handle                  = raft::handle_t(*op_problem.get_handle_ptr());
   rmm::cuda_stream_view barrier_stream = rmm::cuda_stream_per_thread;
-  raft::resource::set_cuda_stream(barrier_handle, barrier_stream);
+  auto barrier_handle                  = raft::handle_t(barrier_stream);
   // Make sure allocations are done on the original stream
   problem.handle_ptr->sync_stream();