data-stream-mgr stream cache not GC'd properly

On bolt80 I ran a reasonably stressful workload (~40 clients) for 16 hours or so.

Over time it seems that the stream_cache_ in data-stream-mgr isn't being cleared as we expect.

After stopping the tests and waiting the cache timeout period (30sec), we would expect the next fragment execution close to expire old cache entries.

On many nodes (I only checked a few of the 72 nodes) the GC only clears a modest number of entries.

E.g. on e1315.halxg.cloudera.com:

I0218 13:10:46.860776 14799 data-stream-mgr.cc:197] Reduced stream ID cache from 40482 items, to 40423, eviction took: 1ms

The cache seems to continue to grow when I run my workload.