Limit file handles used by concurrent readers

[jcsp]

readers_cache should be subject to an overall size limit, and evict readers to make room for new ones if necessary.

Currently it is only implicitly constrainted by its timeout vs. the rate of new readers creation. If a client constructs a series of reads that touches a large number of segments in quick succession, it can grow large (and the underlying reader objects can cause the file handle count to grow large).

JIRA Link: CORE-999