Ability to mlock certain fields from the terms dict [LUCENE-3214]

This is a hacked up prototype! It works but I'm not sure how to get this to a committable point.

The patch invokes mlock() (tested only on Linux), locking pages from the terms dictionary file that hold terms for a specified field. You can only do this with MMapDirectory.

I used this to lock pages for the "id" field in the NRT stress test; it's an alternative to MemoryCodec. But, it requires you set up the OS to allow the app/user to lock pages in RAM.

It works very well in reducing the NRT reopen latency even when large merges are running...

---
Migrated from [LUCENE-3214](https://issues.apache.org/jira/browse/LUCENE-3214) by Michael McCandless (@mikemccand)
Attachments: [LUCENE-3214.patch](https://apache.github.io/lucene-jira-archive/attachments/LUCENE-3214/LUCENE-3214.patch)

Start by reviewing the attached LUCENE-3214.patch and the MMapDirectory and terms-dictionary paths it changes. Check the Linux-only mlock behavior, the required OS permissions, and the NRT stress-test use case; done means the prototype has a committable design with coverage for locking selected field pages and documented limitations.

Written by the indexing model from the issue text.