Warn about loss of precision in FDBSCAN-DenseBox

For some problems and eps, FDBSCAN-DenseBox may suffer from severe loss
of precision when computing cellBox(). This may result in wrong DBSCAN
results.

Initially, the problem was observed on a full NGSIM dataset:
```
$ ./ArborX_DBSCAN.exe --binary \
    --filename /data/data_and_logs/2021_02_20_dbscan_datasets/mustafa_2019/ngsim.arborx \
    --print-dbscan-timers --core-min-size 2 --eps 1.0 --max-num-points 90000 \
    --impl fdbscan-densebox --verify
<...>
Core point is marked as noise: 41888 [-1]
Core point is marked as noise: 17027 [-1]
Core point is marked as noise: 43104 [-1]
```
Tracking down the issue revealed that a point 2279 (part of a dense
cell) was finding point 10369 (not part of a dense cell), but the
reverse was not true (i.e., 10369 was not finding the dense cell that
2279 was in). Studying the dense boxes, I observed that the problematic
box had bounds
  [[6452399.000000,1872182.375000,0.000000],
   [6452399.000000,1872183.000000,0.577350]]
Note the same x coordinate for both min and max corner. The h in this
run was ~0.57. So it completely got lost. Even Y coordinate lost some
precision, though not as much.

src/details/ArborX_DetailsCartesianGrid.hpp

@@ -71,6 +71,9 @@ struct CartesianGrid
   {
     auto min = _bounds.minCorner();
     decltype(min) max;
+
+    // This code may suffer from loss of precision depending on the problem
+    // bounds and h. We try to detect this case in the constructor.
     for (int d = 0; d < DIM; ++d)
     {
       auto i = cell_index % _n[d];
@@ -118,6 +121,23 @@ struct CartesianGrid
       m /= _n[d - 1];
       ARBORX_ASSERT(_n[d] < m);
     }
+
+    // Catch a potential loss of precision that may happen in cellBox() and can
+    // lead to wrong results.
+    //
+    // The machine precision by itself is not sufficient. In some experiments
+    // run with a full NGSIM datasets, values below 3 could still produce wrong
+    // results. This may still not be conservative enough, but all runs passed
+    // verification when this warning was not triggered.
+    float constexpr eps = 5 * std::numeric_limits<float>::epsilon();
+    for (int d = 0; d < DIM; ++d)
+    {
+      if (std::abs(_h[d] / min_corner[d]) < eps)
+        throw std::runtime_error(
+            "ArborX exception: FDBSCAN-DenseBox algorithm will experience loss "
+            "of precision, undetectably producing wrong results. Please switch "
+            "to using FDBSCAN.");
+    }
   }
 
   Box _bounds;