Implement column sampler in CUDA. (#9785)

- CUDA implementation.
- Extract the broadcasting logic, we will need the context parameter after revamping the collective implementation.
- Some changes to the event loop for fixing a deadlock in CI.
- Move argsort into algorithms.cuh, add support for cuda stream.

src/tree/gpu_hist/evaluator.cu

@@ -72,7 +72,7 @@ common::Span<bst_feature_t const> GPUHistEvaluator::SortHistogram(
     TreeEvaluator::SplitEvaluator<GPUTrainingParam> evaluator) {
   dh::XGBCachingDeviceAllocator<char> alloc;
   auto sorted_idx = this->SortedIdx(d_inputs.size(), shared_inputs.feature_values.size());
-  dh::Iota(sorted_idx);
+  dh::Iota(sorted_idx, dh::DefaultStream());
   auto data = this->SortInput(d_inputs.size(), shared_inputs.feature_values.size());
   auto it = thrust::make_counting_iterator(0u);
   auto d_feature_idx = dh::ToSpan(feature_idx_);