Critical bug fix, performance improvements, less code-gen, and more!

This release includes a critical fix to avoid double reading first block in split in block-format.

It also uses dynamic protobuf instead of code-generating a protobuf for the block-format, has some performance improvements in base64 codepaths, and more!

Here's the full change log:
Issue 429 Avoid double reading first block in split in block-format (Raghu Angadi)
Issue 421. Expose FileDescriptor and FieldDescriptor (Brian Ramos)
Issue 423. Don't copy the array before handing it for base64 decode (Ian O'Connell)
Issue 422. Pulls in the source of a BSD licenced base64 implementation that is 5x faster than the Apache one for our usage (Ian O'Connell)
Issue 418. Use dynamic protobufs (Remove protobufs) (Raghu Angadi)
Issue 417. A Cascading scheme for combining intermediate sequence files (Akihiro Matsukawa)
Issue 414. Fix typo in docs (thrift, not thrist) (gstaubli)
Issue 413. Trivial Javadocs for LuceneIndexInputFormat (Lewis John McGibbney)
Issue 412. Adding support for Map, Sets and Lists to ThriftToDynamicProto (Brian Ramos)
Issue 411. Fix NPE in CompositeRecordReader due to improper delegate initialization (Jonathan Coveney)
Issue 409. Gzip objects before storing them in the job conf (Alex Levenson)
Issue 407. Make dependencies expicit in Readme quickstart (fixes #406) (Lewis John McGibbney)
Issue 405. Fix bug in CompositeRecordReader (Jonathan Coveney)
Issue 403. Refactor CompositeRecordReader to only make a record reader when necessary (Jonathan Coveney)
Issue 398. Add CombineFileInputFormat support (esp. for lzo) (Jonathan Coveney)