Fix hanging pytests (#1924)

This PR attempts to fix the hanging pytest issues we are seeing in CI.
From my investigations there are two issues.

One is the `test_cagra_ace_tiny_memory_limit_triggers_disk_mode` and `test_hnsw_ace_tiny_memory_limit_triggers_disk_mode` tests which I have seen locally to hang. ~~At the moment, I am skipping these pytests in CI.~~
**Update**: After testing here 20 times https://github.com/rapidsai/cuvs/actions/runs/23182916436?pr=1893, I don't think we need to skip these tests.

Second is a memory out of bounds error in `device_matrix_view_from_host` , first revealed by compute-sanitizer. This is likely a HMM issue. I have added more specific checks in the constructor so a copy is used when necessary.

See this CI run from my test PR where I run 10 times to ensure that CI works with these fixes.
https://github.com/rapidsai/cuvs/actions/runs/23131230353
I also reran CI on that test PR again. So all the pytests were run 20 times.

Authors:
  - Anupam (https://github.com/aamijar)

Approvers:
  - Tarang Jain (https://github.com/tarang-jain)

URL: #1924

Author: aamijar

cpp/src/neighbors/detail/cagra/utils.hpp

@@ -174,8 +174,10 @@ class device_matrix_view_from_host {
   {
     cudaPointerAttributes attr;
     RAFT_CUDA_TRY(cudaPointerGetAttributes(&attr, host_view.data_handle()));
-    device_ptr = reinterpret_cast<T*>(attr.devicePointer);
-    if (device_ptr == NULL) {
+    device_ptr      = reinterpret_cast<T*>(attr.devicePointer);
+    bool needs_copy = (device_ptr == NULL) ||
+                      (attr.type != cudaMemoryTypeDevice && attr.type != cudaMemoryTypeManaged);
+    if (needs_copy) {
       // allocate memory and copy over
       // NB: We use the temporary "large" workspace resource here; this structure is supposed to
       // live on stack and not returned to a user.