Think through implications of blobs for buffering

To be efficient, the blobs output should be buffering at least 1Mb of data, (but local files probably want a smaller size).  However we don't want multiple buffering layers, and it would be better if the compression layer (and other items e.g. index building) could write directly into the output buffer.

I think any changes are best tied to refactoring the interfaces to use a streaming output interface more of the time.