CAGRA graph optimizer: clamp rev_graph_count (#1987)

This PR fixes a potential out of bounds access. 

While building the reverse graph, we count the number of potential edges in `rev_graph_count`

https://github.com/rapidsai/raft/blob/047bfb2a7a24a97a4bece0b553521533868e7889/cpp/include/raft/neighbors/detail/cagra/graph_core.cuh#L204-L205

While we store only max `output_graph_degree` neighbors, the `rev_graph_count` can be larger than `output_graph_degree`.  The loop that uses `rev_graph` is updated to take this limit into account.

Authors:
  - Tamas Bela Feher (https://github.com/tfeher)

Approvers:
  - tsuki (https://github.com/enp1s0)
  - Corey J. Nolet (https://github.com/cjnolet)

URL: #1987

cpp/include/raft/neighbors/detail/cagra/graph_core.cuh

@@ -528,8 +528,9 @@ void optimize(raft::resources const& res,
     constexpr int _omp_chunk = 1024;
 #pragma omp parallel for schedule(dynamic, _omp_chunk)
     for (uint64_t j = 0; j < graph_size; j++) {
-      for (uint64_t _k = 0; _k < rev_graph_count.data_handle()[j]; _k++) {
-        uint64_t k = rev_graph_count.data_handle()[j] - 1 - _k;
+      uint64_t k = std::min(rev_graph_count.data_handle()[j], output_graph_degree);
+      while (k) {
+        k--;
         uint64_t i = rev_graph.data_handle()[k + (output_graph_degree * j)];
 
         uint64_t pos =