Incremental offline decoding/alignment API needed

[dhdaines]

For force alignment we are not actually doing live recognition, but the API inherited from PocketSphinx still assumes this. We need a better one, which:

• Assumes that the entire input is available (though perhaps not all in memory)
• Does CMN on a sufficient quantity of input to get good results
• Emits progress events (which can be wrapped in an Observable)
• Breaks the computation into small enough pieces to be run asynchronously in the main thread

[dhdaines]

In fact this may need to potentially support quite large input files...