zpool iostat -ow object store latency buckets off by one (openzfs#823)

In `zpool iostat -ow`, the `object_store` latencies are being counted in
one higher bucket than they should be (e.g. 512us instead of 256us),
making it hard to compared to the `object_agent` latencies.  The bucket
index is somewhat confusing due to the kernel's definition of
`L_HISTO()`.

This commit fixes the issue.

cmd/zfs_object_agent/zettaobject/src/access_stats.rs

@@ -122,7 +122,7 @@ impl<'a> OpInProgress<'a> {
 
         // This bucket mapping is equivalent to L_HISTO() macro in zfs.h
         let latency_bucket = std::cmp::min(
-            latency.next_power_of_two().trailing_zeros(),
+            latency.next_power_of_two().trailing_zeros() - 1,
             LATENCY_HISTOGRAM_BUCKETS - 1,
         ) as usize;