Remove dependency on Boost.IOStreams

[arahlin]

Implement file IO functions using STL constructs.

A set of stream buffer classes are defined in the private streams.h header. API functions in the public dataio.h header are significantly reduced, relying on the stream buffers to provide the backend implementation for flush, seek, tell, etc. The callback registration API is used to ensure destruction of instantiated stream buffers.

The compression implementation here is API-compatible with the boost variant, verified by writing files with this branch and reading files with the master branch, and vice versa. These compression streambuf classes could straightforwardly be used in other applications, for example to implement the hybrid bzip2/flac compression scheme used in G3SuperTimestreams.

Seek/tell efficiency of the reader is maintained, although slightly reimplemented here with a caching counter.

New features:

• Both reading and writing streams use a (tunable) 1MB buffer, for compressed, uncompressed, local and remote streams.
• Update ARCFileReader API to expose the timeout and buffersize istream options.

Fixes #185.

[arahlin]

Performance testing showed that the slow tellg issue cropped up again, so I implemented a byte counter that seems to work. Otherwise read/write speeds are the same as the boost implementation.

[nwhitehorn]

This looks good to me from an architectural perspective. I'm not 100% I understand the rationale for the file-extension arguments, but that's a minor quibble. Thanks!

[cnweaver]

As a general correctness issue, I don't see any calls to ios::xalloc, but it is a precondition that ios::pword can only be called with index values previously returned by xalloc. (I think the point here is that any user code is allowed to use pword and iword to attach things to ios_base, so each distinct use should call xalloc to ensure that it is using a unique index to do so. Hard-coding it to zero everywhere is unsafe, because it can conflict with any other (rare) code using this feature. It looks like there should be central bit of code that calls xalloc once, either the first time the index is needed, or on startup, and caches the value it gets in a variable which every function calling pword should read to get the index to use.