First this is great. We've had requests to read Sequence files. Being able to write them also would be great.
A few thoughts:
1) This should not extend UTF8StorageConverter. This loader will be returning actual data types, not bytes that need to be interpreted. I would think instead that it should implement the bytesToX() methods itself and just throw an exception saying it didn't expect to do any conversion.
2) The getSampledTuple looks fine if skip is handling getting the stream to the point that reading the next tuple is viable.
3) In the bindTo call, where you obtain the key and value by reflection, should there be a try/catch block there in case the cast to Writable fails? In the same way, in describe schema you're asking how to suppress warnings from the cast in reader.getKeyClass(). But don't you want to check that what you got really is a writable, since there is no guarantee?