Reset CUDA error after cudaFuncSetAttribute (#504)

This PR fixes a bug which occures when `cudaFuncSetAttribute` is used to
conditionally check if there's enough shmem available to fit the HLL
sketch in. If it doesn't fit then the CUDA error returned by the
function is sticky and may resurface in downstream calls, e.g., Thrust.
The fix consists of flushing the CUDA error after the call to
`cudaFuncSetAttribute`.

Big thanks to @gevtushenko for tracking down this bug.

Author: sleeepyjack

include/cuco/detail/hyperloglog/hyperloglog_ref.cuh

@@ -550,9 +550,12 @@ class hyperloglog_ref {
   [[nodiscard]] __host__ constexpr bool try_reserve_shmem(Kernel kernel,
                                                           int shmem_bytes) const noexcept
   {
-    return cudaSuccess == cudaFuncSetAttribute(reinterpret_cast<void const*>(kernel),
-                                               cudaFuncAttributeMaxDynamicSharedMemorySize,
-                                               shmem_bytes);
+    bool const ret =
+      cudaSuccess == cudaFuncSetAttribute(reinterpret_cast<void const*>(kernel),
+                                          cudaFuncAttributeMaxDynamicSharedMemorySize,
+                                          shmem_bytes);
+    cudaGetLastError();  // flush CUDA error
+    return ret;
   }
 
   hasher hash_;                            ///< Hash function used to hash items